Spike cameras represent a novel class of neuromorphic imaging devices that capture visual scenes as binary spike streams through temporal integration of light intensity. While offering exceptional dynamic range and energy efficiency, they generate massive binary data volumes that require efficient compression. Existing codecs fail to exploit the unique structure of spike data and lack support for arbitrary temporal resolution. We propose LSE-Codec, a neural lossless compression framework specifically designed for spike cameras that operates independently of frame-rate, as shown in Fig. 1. Our approach introduces Local Spike Embedding (LSE) to reorganize sparse binary spike patterns into compact 8 -bit symbols while preserving spatial structure, followed by a hierarchical autoencoder with autoregressive entropy modeling to predict spike distributions. Evaluated on different datasets under aligned test conditions, our method achieves state-of-the-art performance, reducing bit rates by over 11% compared to JPEG-XL (best-effort mode). This work bridges a critical gap in spike vision systems with frame-rate agnostic lossless coding, enabling efficient storage and transmission of spike data without sacrificing fidelity.

Recent advances in cross-modal compression(CMC) have opened new horizons for perceptual image coding at ultra-low bitrates (below 0.1 bpp) within a generative compression paradigm, but reconstruction fidelity is often compromised, yielding visually plausible yet semantically inconsistent reconstructions. While prompt engineering with contextual optimization has been extensively explored in generative models, its potential for controlling perception-fidelity trade-offs in image compression remains largely under-explored. To address these challenges, we propose PO-CMC, a novel diffusion-based cross-modal image compression approach that introduces contextual prompt optimization to achieve efficient and perceptually faithful reconstruction. The proposed method comprises three synergistic components: an optimized image codec that produces a compact structural prior, a contextual prompt module that adaptively encodes semantic cues into compact textual embeddings, and a diffusion-based decoder that fuses the structural and semantic priors to reconstruct high-fidelity images. Extensive experiments show that PO-CMC achieves superior perceptual quality while maintaining comparable reconstruction fidelity, yielding an average BD-rate saving of 72.5 % and 79.8 % over VVC at equivalent LPIPS and DISTS levels, respectively.

The recent development of feedforward 3D Gaussian Splatting (3DGS) presents a new paradigm to reconstruct 3D scenes. Using neural networks trained on large-scale multi-view datasets, it can directly infer 3DGS representations from sparse input views. Although the feedforward approach achieves high reconstruction speed, it still suffers from the substantial storage cost of 3D Gaussians. Existing 3DGS compression methods relying on scene-wise optimization are not applicable due to architectural incompatibilities. To overcome this limitation, we propose TinySplat, a complete feedforward approach for generating compact 3D scene representations. Built upon standard feedforward 3DGS methods, TinySplat integrates a training-free compression framework that systematically eliminates key sources of redundancy. Specifically, we introduce View-Projection Transformation (VPT) to reduce geometric redundancy by projecting geometric parameters into a more compact space. We further present Visibility-Aware Basis Reduction (VABR), which mitigates perceptual redundancy by aligning feature energy along dominant viewing directions via basis transformation. Lastly, spatial redundancy is addressed through an off-the-shelf video codec. Comprehensive experimental results on multiple benchmark datasets demonstrate that TinySplat achieves over 100× compression for 3D Gaussian data generated by feedforward methods. Compared to the state-of-the-art compression approach, we achieve comparable quality with only 6% of the storage size. Meanwhile, our compression framework requires only 25% of the encoding time and 1% of the decoding time.

With advancement of deep space exploration, vast amounts of image data must be transmitted back to Earth for scientific research. Current deep-space image codecs rely on conventional progressive coding algorithms, but offer limited compression performance. Despite great success on natural images achieved by learning-based compression methods, the high computational complexity restrains their application in the deep space missions and they are unable to cope with packet loss caused by severe noise interference during transmission. Motivated by the urgent need and technical challenges, we take Mars as a representative case and propose a novel image compression and transmission framework that innovatively incorporates learning-based strategies to deliver both low-complexity and error-resilient source coding. To adapt learning-based methods to the stringent constraints and high packet loss of deep space environment, we first establish a new Martian image dataset with high resolution and diversity, and analyze its characteristics to guide the network design. With heterogeneous textures yet synergistic structures as well as higher inter-channel similarity in the feature domain revealed for the Martian images, we develop a Martian Vision Adaptive Transformation Module (MVATM) with efficient low-complexity compression. Furthermore, unlike conventional one-stage training, a novel two-stage training strategy with Joint Channel Training (JCT) is proposed to enhance error resilience. Experimental results and hardware deployment strongly validate that our method achieves a better rate-distortion-complexity (RDC) trade-offs than other advanced learning-based models and significantly outperforms conventional methods in deep space simulation test. Also, the technical strategies proposed herein can offer methodological insights for deep space and other resource-constrained fields.

This paper introduces the Spatio-Temporal Adaptive Context Optimization (STACO) method, which enhances the quality of contextual prediction across various resolutions, essential for subsequent compression. The STACO takes predicted contexts$C_t^{\{1,2,3\}}$as input and improves their quality by aligning them better with decoded features$f_{t}$, thus boosting coding efficiency. The STACO comprises Quality Perception Units and Consistency Synergy Modules, arranged in a hierarchical stacked architecture. This multi-scale design enables simultaneous processing of contexts at different spatial resolutions and facilitates information exchange through upsampling and downsampling. Enhanced contexts$\tilde{C}_{t}^{\{1,2,3\}}$are output after passing through residual connections, ensuring better alignment with reconstructed features. Using VTM-11.0 as anchor, the STACO significantly improves compression efficiency on common test condition (CTC) in HEVC, achieving an average BD-rate reduction of 17.99% for PSNR and 43.84% for MS-SSIM. By incorporating spatial quality mapping and temporal propagation, STACO offers a significant advancement in video compression.

For decades, the Bjøntegaard Delta (BD) has been the metric for evaluating codec Rate-Distortion (R-D) performance. Yet, in most studies, BD is determined using just 4–5 R-D data points, could this be sufficient? As codecs and quality metrics advance, does the conventional BD estimation still hold up? Crucially, are the performance improvements of new codecs and tools genuine, or merely artifacts of estimation flaws? We address these concerns by reevaluating BD estimation. We have established a large-scale, high-precision R-D dataset to verify the accuracy of existing BD estimation algorithms. Moreover, we propose a robust method for high-precision BD estimation across diverse compression scenarios, enhanced by a reliability assessment to determine the probability distribution of BD values from R-D sample points. This approach both assesses the reliability of BD calculations and serves as a precise BD estimator. Our method's validity is confirmed through extensive testing on a dataset we constructed. Our findings advocate for the adoption of rigorous R-D sampling and reliability metrics in future compression research to ensure the validity and reliability of results. Our code and additional experimental details are publicly accessible at https://github.com/fgvfgfg564/BDCI.

This paper propose a Image Coding for Machine (ICM) framework with Visual-Language Mimic Feature Learning (VLM-ICM). VLM-ICM decouples the position and semantic information into language modality and extracts universal features from the input image. Language, inherently more semantically compact, helps reduce the bitrate. Meanwhile, the universal features in VLM-ICM, guided by the language at the decoder side, allow for flexible domain adaptation, thereby enhancing versatility and practicality.

Neural Video Compression (NVC) has advanced significantly in recent years, with improvements in inter prediction techniques. In inter prediction, most NVC approaches utilize pixel information or temporal features from neighboring frames as reference information, while using optical flow to represent motion information. In this paper, we introduce an innovative and efficient method for Dynamic Temporal Reference Aggregation (DTRA). The proposed DTRA consists of two components: Temporal Information Compensation (TIC) and Feature Level Motion Information Enhancement (MIE). The TIC module generates compensation information by leveraging long-term temporal information from the decoding buffer, enriching the semantic content of the reference features and enhancing their texture details. The MIE module refines the motion features at the encoder side and divides the motion information into multiple groups for diverse motion alignment at the decoder side, thereby improving the motion compensation. Extensive experiments demonstrate the effectiveness of the proposed method, achieving an average bitrate savings of 9.67% compared to state-of-the-art (SOTA) approaches.

Screen content is an important category of video. The statistical distribution of pixels in screen content exhibits substantial differences compared to camera-captured video, leading to different compression needs and challenges. Hitherto, most Screen Content Coding (SCC) tools are designed for block-based hybrid coding frameworks, which may not be suitable for emerging wavelet-based and learning-based coding frameworks. Consequently, a plug-and-play SCC tool independent of coding frameworks is lacking in the current video coding landscape. In this paper, an out-loop coding method, Frequency-based Adaptive Pixel Correction (FAPC), is proposed to improve the SCC performance for arbitrary codecs. First, a True Color Value (TCV) table is established based on the most frequently occurring pixel values. Then, the reconstructed pixels are corrected according to the TCV table. To realize precise pixel correction, an adaptive threshold derivation method is meticulously designed to control the pixel correction process. Furthermore, an inheritance coding strategy is proposed to reduce the overhead of parameter transmission. The proposed method has been integrated into three different coding frameworks. Simulation results demonstrate that the proposed method can achieve 10.66%, 3.55% and 10.67% luma component BD-BR gains on the three coding frameworks, respectively. These results prove the superiority and universality of the proposed method.

Image Compression for Machines (ICM) aims to compress images for machine vision tasks, while current methods mostly focus on the demands for high-level tasks. However, the quality of original images is usually not guaranteed in the real world, leading to even worse downstream task performance after compression. Thus, lowlevel (LL) restoration tasks should also be considered in ICM. In this paper, we propose the first ICM framework for LL machine vision tasks, namely LL-ICM, which optimizes the compression and LL processing performance simultaneously. Moreover, LL-ICM leverages large vision-language model (VLM) to solve different LL task within a single model, which is particularly useful when the distortion type of the original image is uncertain. As illustrated in Fig. 1(a), LL-ICM consists of a neural image codec and a VLM-based LL processing module. Given an original image with distortions, LL-ICM firstly compress it as$\hat{\mathbf{X}}$. Then, we extract a generalized feature F from$\hat{\mathbf{X}}$, which is then encoded as two representations, distortion type$\varphi$and caption$\sigma$. After that, the LL processing module receives$\hat{\mathbf{X}}$and its representations to generate the restored version of$\hat{\mathbf{X}}$, i.e.,$\hat{\mathbf{X}}_{\mathbf{H}}$.

Logarithmic binary arithmetic coding (LBAC) is a entropy coding method first introduced in AVS2 and now used in the third generation of audio video coding standard (AVS3). While LBAC provides high coding efficiency, the data dependencies in AVS3 make it challenging to parallelize parsing syntax element. To enhance the throughput of AVS3 LBAC entropy decoding engine, a sophisticated hardware implementation architecture is meticulously designed in the paper. The proposed method decouples the bitstream decoding process into four distinct modules. To ensure the decoding efficiency and performance, an advanced parallel pipeline is carefully designed. Furthermore, we develop sub-branch state prediction mechanism and grouped decoding method to further improve the decoding efficiency. The proposed AVS3 entropy decoder has been implemented on the S10 FPGA System Board, and experimental results show that the proposed method could achieve real-time and seamless decoding efficiency for 8K AVS3 bitstreams, even at bitrates as high as 207.1Mbps.

360-degree video is a panoramic video technology designed to offer audience an immersive visual experience. In Motion Constrained Tile Set (MCTS)-based streaming schemes, the server only updates Field-Of-View (FOV) coordinates when codec generating keyframes, thus the keyframe interval significantly impacts FOV updates and interactivity. However, reducing the keyframe interval poses greater challenges for network transmission and encoding/decoding overhead. To address these issues, we design and implement AKI360, a 360-degree video streaming system with an Adaptive Keyframe Interval (AKI) mechanism along with Quantization Parameter (QP) adjustments. Extensive experiments show that AKI360 reduces the motion stall duration by 31% to 74% and FOV update interval by 78% to 86% comparing with the real-time approaches, and eliminates the blank area in the best-case comparing with the state-of-the-art pre-encoded approach. Meanwhile, AKI360 improves the video quality by up to 0.78 decrease in NIQE while maintaining comparable frame rate.

Learned Image Compression (LIC) has rapidly evolved and recently surpassed traditional methods in Rate-Distortion (R-D) performance. However, most LIC approaches improve network architectures while increasing computational overhead and overlooking using optimizers tailored specifically for LIC. This paper proposes a Dynamic Momentum-Smoothing Optimizer (DMSO) tailored for LIC to achieve faster convergence and better R-D performance. Specifically, DMSO leverages historical gradient information to smooth the optimization process dynamically, thereby reducing in-stability from gradient oscillations. Furthermore, DMSO introduces a novel Enhanced Second-order Momentum mechanism to mitigate cumulative noise and align momentum updates more closely with the true gradient. Experimental results demonstrate that DMSO operates as a universal optimization plugin for LIC methods, achieving faster and more stable convergence while improving R-D performance to varying degrees without additional parameter count or computational cost.

Video compression is a fundamental topic in the visual intelligence, bridging visual signal sensing/capturing and high-level visual analytics. The broad success of artificial intelligence (AI) technology has enriched the horizon of video compression into novel paradigms by leveraging end-to-end optimized neural models. In this survey, we first provide a comprehensive and systematic overview of recent literature on end-to-end optimized learned video coding, covering the spectrum of pioneering efforts in both uni-directional and bi-directional prediction based compression model designation. We further delve into the optimization techniques employed in learned video compression (LVC), emphasizing their technical innovations, advantages. Some standardization progress is also reported. Furthermore, we investigate the system design and hardware implementation challenges of the LVC inclusively. Finally, we present the extensive simulation results to demonstrate the superior compression performance of LVC models, addressing the question that why learned codecs and AI-based video technology would have with broad impact on future visual intelligence research.

Neural Network (NN)-based video coding technologies have emerged as a promising alternative to traditional methods, demonstrating significant advantages amidst the rapid advancements in video coding technology. This paper presents a hybrid video coding method based on the Enhanced Compression Model (ECM) developed by the Joint Video Exploration Team (JVET). We integrate two NN-based coding tools into the framework. Specifically, the proposed NN-based Intra Prediction (NNIP) method effectively models the nonlinear relationship between neighboring contextual information and the block to be predicted. The NN-based In-Loop Filtering (NNILF) method adaptively filters the luminance and chrominance components across various quality levels. Experimental results show that the NNIP and NNILF methods achieve 0.56% and 4.14% BD-rate savings for YCbCr components under the All Intra (AI) configuration compared to ECM-14.0. Under the Random Access (RA) configuration, the proposed method can achieve a 2.41% BD-rate saving for YCbCr components.

Recent advances in 3D Gaussian Splatting (3DGS) have enabled fast, photorealistic rendering of dynamic 3D scenes, showing strong potential in immersive communication. However, in digital human encoding and transmission, the compression methods based on general 3DGS representations are limited by the lack of human priors, resulting in suboptimal bitrate efficiency and reconstruction quality at the decoder side, which hinders their application in streamable 3D avatar systems. We propose HGC-Avatar, a novel Hierarchical Gaussian Compression framework designed for efficient transmission and high-quality rendering of dynamic avatars. Our method disentangles the Gaussian representation into a structural layer, which maps poses to Gaussians via a StyleUNet-based generator, and a motion layer, which leverages the SMPL-X model to represent temporal pose variations compactly and semantically. This hierarchical design supports layer-wise compression, progressive decoding, and controllable rendering from diverse pose inputs such as video sequences or text. Since people are most concerned with facial realism, we incorporate a facial attention mechanism during StyleUNet training to preserve identity and expression details under low-bitrate constraints. Experimental results demonstrate that HGC-Avatar provides a streamable solution for rapid 3D avatar rendering, while significantly outperforming prior methods in both visual quality and compression efficiency.

Learned Image Compression (LIC) has attracted considerable attention due to their outstanding rate-distortion (R-D) performance and flexibility. However, the substantial computational cost poses challenges for practical deployment. The issue of feature redundancy in LIC is rarely addressed. Our findings indicate that many features within the LIC backbone network exhibit similarities. This paper introduces ShiftLIC, a novel and efficient LIC framework that employs parameter-free shift operations to replace large-kernel convolutions, significantly reducing the model’s computational burden and parameter count. Specifically, we propose the Spatial Shift Block (SSB), which combines shift operations with small-kernel convolutions to replace large-kernel. This approach maintains feature extraction efficiency while reducing both computational complexity and model size. To further enhance the representation capability in the channel dimension, we propose a channel attention module based on recursive feature fusion. This module enhances feature interaction while minimizing computational overhead. Additionally, we introduce an improved entropy model integrated with the SSB module, making the entropy estimation process more lightweight and thereby comprehensively reducing computational costs. Experimental results demonstrate that ShiftLIC outperforms leading compression methods, such as VVC Intra and GMM, in terms of computational cost, parameter count, and decoding latency. Additionally, ShiftLIC sets a new SOTA benchmark with a BD-rate gain per MACs/pixel of −102.6%, showcasing its potential for practical deployment in resource-constrained environments. The code is released athttps://github.com/baoyu2020/ShiftLIC.

Recently, the bio-inspired spike camera with continuous motion recording capability has attracted tremendous attention due to its ultra high temporal resolution imaging characteristic. Such imaging feature results in huge data storage and transmission burden compared to that of traditional camera, raising severe challenge and imminent necessity in compression for spike camera captured content. Existing lossy data compression methods could not be applied for compressing spike streams efficiently due to integrate-and-fire characteristic and binarized data structure. Considering the imaging principle and information fidelity of spike cameras, we propose a novel Reconstruction-based Contextual Spike Compression (RCSC) framework, which contains scene reconstruction, contextual image compression and spike generation. To our knowledge, it is the first learning-based model for efficient and robust spike stream compression with informative fidelity. Extensive experimental results show that our model outperforms the state-of-the-art conventional codec VVC intra by 6.14% and surpasses the state-of-the-art learned codec by 2.53% in BD-rate reduction, establishing a strong baseline for spike compression.

End-to-end optimized Learned Image Compression (LIC) has demonstrated remarkable performance in terms of Rate-Distortion (R-D) efficiency. However, the R-D characteristics of LIC codecs remain underexplored. Previous research has attempted to investigate the R-D behavior through numerical and statistical approaches, but these methods often provide only empirical results, lacking theoretical insights. In this work, we introduce a novel methodology for studying the R-D characteristics of LIC. By rethinking the LIC paradigm from a fresh perspective, we propose a plug-and-play module, the Latent-domain Auto-Encoder (LAE). This innovative approach not only naturally leads to Variable Bit-Rate (VBR) compression, but also allows for a theoretical modeling of the R-D behavior of LIC codecs. Our findings reveal that the bit-rate is the logarithmic sum of the neuronsnλin our designed network’s last layer, plus a constantCintroduced by image content, formally expressed asRλ= Σ lognλ+C. This insight is pivotal, as it underscores how the bit-rate can be systematically derived from the latent representations. Further analysis demonstrates that our proposed R-λ model enables effective rate control for learned image codecs, enhancing their adaptability and accuracy. Experimental results validate that our VBR method surpasses fixed-rate coding by 2.9% in terms of BD-rate. Additionally, the proposed R-λ model exhibits superior rate control performance, suggesting that it not only elucidates the underlying R-D characteristics of LIC but also significantly enhances its practical deployment in real-world applications.

Recent studies on Video Coding for Machine (VCM) have achieved remarkable results. However, in practical visual-analytic applications, except for transmitting the visual feature for machine analysis, video textures are also mandatory for human monitoring and decision-making. To this end, this paper proposes a human-machine friendly video compression scheme (HMFVC) which can satisfy both human viewing and machine analysis well. First, we propose a learned semantic representation (LSR) method to extract semantic information between temporal neighboring frames. LSR could be utilized in signal reconstruction for human viewing and visual analysis for machine understanding. Second, given the proposed LSR, we design an end-to-end optimized video compression framework to jointly optimize the visual quality for human perception, analysis accuracy for machines, and compression efficiency as well. Finally, an HMFVC codec is developed, which can achieve higher action recognition accuracy and better reconstruction quality than the traditional codecs and learned video compression approaches. Specifically, HMFVC saves 77% bitrate to achieve the same analysis performance with the original videos compared to x265. To our knowledge, HMFVC is the first end-to-end optimized video compression scheme to serve both humans and machines. It is a promising framework for human-machine friendly video compression approaches.

In this age of information, images are a critical medium for storing and transmitting information. With the rapid growth of image data amount, visual compression and visual data perception are two important research topics attracting a lot of attention. However, those two topics are rarely discussed together and follow separate research paths. Due to the compact compressed domain representation offered by learning-based image compression methods, there exists possibility to have one stream targeting both efficient data storage and compression, and machine perception tasks. In this paper, we propose a layered generative facial image compression model achieving high human vision-oriented image reconstructed quality, even at extreme compression ratios. To obtain analysis efficiency and flexibility, a task-agnostic learning-based compression model is proposed, which effectively supports various compressed domain-based analytical tasks while preserving outstanding reconstructed perceptual quality, compared with traditional and learning-based codecs. In addition, joint optimization schedule is adopted to acquire best balance point among compression ratio, reconstructed image quality, and downstream perception performance. Experimental results verify that our proposed compressed domain-based multi-task analysis method can achieve comparable analysis results against the RGB image-based methods with up to 99.6% bit rate saving (i.e., compared with taking original RGB image as the analysis model input). The practical ability of our model is further justified from model size and information fidelity aspects.

Optical remote sensing images (ORSI) feature unique scenes and complex imaging conditions. Specifically, they exhibit substantial variations in object scale, quantity, structure, and distribution. Consequently, salient object detection in ORSI (ORSI-SOD) is pivotal in ORSI content perception and understanding. Additionally, the limitations of the single modality impede the advancement of ORSI-SOD. To tackle these issues, we propose a Depth-guided and Iterative Refinement Network (DINet) for ORSI-SOD. By incorporating depth information as auxiliary cues, we introduce a multi-modal strategy for ORSI-SOD, resulting in improved accuracy in the localization and segmentation of salient objects. To address the variability of salient objects, we design an Aggregation Perception Enhancement (APE) Module. This module integrates complementary cues from cross-modal features using multi-dimensional attention mechanisms. By fostering cross-modal interactions, the APE module effectively preserves both detail and spatial location information. Furthermore, we propose an Iterative Guidance Refinement Decoder to handle boundary uncertainty. The decoder uses initial predictions to guide the decoding phase and iteratively refine results. Simultaneously, it minimizes noise from depth cues, yielding predictions with more accurate boundaries. Experimental comparisons with 22 state-of-the-art methods show that DINet exhibits superior performance while maintaining lightweight (11.98M) and real-time (55FPS) capabilities.

Spike cameras have shown great potential in capturing ultra-high-speed motion scenes by mimicking the retinal fovea's function, especially addressing the challenges of full-time imaging and high dynamic range in an energy-efficient fashion. Leveraging spike emission mechanisms, these cameras achieve extraordinary temporal resolutions in terms of thousands of frames per second, far surpassing traditional imaging devices. However, the resulting data, characterized by its large scale and sufficient temporal imaging nature, poses significant challenges for storage and transmission. In this paper, we propose an advanced lossless compression model for spike data via constructing a novel spike data representation scheme. We first introduce an efficient short-term aggregation method for spike sequences, paired with an intensity remapping technique to mitigate the effects of noise inherent in the spike sampling approach. In addition, we design and propose the Categorical Logit-based Entropy Model (CLEM) by quantitatively and precisely measuring the required code length of the underlying representation to generate an implicit representation that models the unique statistical distribution of spike data. We leverage these findings to introduce a novel learned lossless spike compression model that significantly reduces the data rate while preserving full data fidelity. Extensive experimental results on PKU-Spike-Recon and more real-world spike datasets demonstrate that our approach achieves state-of-the-art (SOTA) performance, with competitive computational complexity. The proposed method illuminates a new path towards lossless compression without encoding the prediction residual for spike data coding.

The advent of Neuromorphic spike cameras has garnered significant attention for their ability to capture continuous motion with unparalleled temporal resolution. However, this imaging attribute necessitates considerable resources for binary spike data storage and transmission. In light of compression and spike-driven intelligent applications, we present the notion of Spike Coding for Intelligence (SCI), wherein spike sequences are compressed and optimized for both bit-rate and task performance. Drawing inspiration from the mammalian vision system, we propose a dual-pathway architecture for separate processing of spatial semantics and motion information, which is then merged to produce features for compression. A refinement scheme is also introduced to ensure consistency between decoded features and motion vectors. We further propose a temporal regression approach that integrates various motion dynamics, capitalizing on the advancements in warping and deformation simultaneously. Comprehensive experiments demonstrate our scheme achieves state-of-the-art (SOTA) performance for spike compression and analysis. We achieve an average 17.25% BD-rate reduction compared to SOTA codecs and a 4.3% accuracy improvement over SpiReco for spike-based classification, with 88.26% complexity reduction and 42.41% inference time saving on the encoding side.

Knowledge is an abstraction of factual principles of the physical world. Large foundation models encapsulate extensive multimodal knowledge into the parameters and thus invoke machine intelligence on various tasks. How to invoke the knowledge in these models to facilitate image compression lacks in-depth exploration. In this work, we aim to harness multimodal knowledge into ultra-low bitrate compression and propose Multimodal Knowledge-aware Image Compression (MKIC). Our key insight is that under the context of ultra-low bitrate compression, where the encoded representation is too sparse to represent enough information of the input signal, knowledge from the physical world is required to be incorporated into the compression. Thus, more shared patterns can be stored in the model together with sparse unique features also embedded into the bitstream. In light of two kinds of knowledge, namely natural visual knowledge and human language knowledge, we propose a novel Alternating Rate-Distortion Optimization to enhance the accuracy and compactness of global semantic text representation extraction, extract the local feature map that captures visual details, and integrate these multimodal representations into a large generative foundation model to achieve high-quality reconstruction. The proposed method relights the path of learned image coding, leveraging decoupled knowledge from large foundation models. Extensive experiments show that our proposed method achieves superior comprehensive performance compared to various methods and shows great potential for ultra-low bitrate image compression.

Learned Image Compression (LIC) has experienced rapid growth with the emergence of diverse frameworks. However, the variability in model design and training datasets poses a challenge for the universal application of a single coding model. To address this problem, this paper introduces a pioneering multi-model image coding framework that integrates various image codecs to overcome these limitations. By dynamically allocating codecs to different image regions, our framework optimizes reconstruction quality within the constraints of limited bitrate and decoding time, offering a high-performance, ubiquitous solution for the rate-distortion-complexity trade-off. Our framework features a detailed codec assignment algorithm based on the Simulated Annealing (SA) method, selected for its proven efficacy in managing the discrete and intricate nature of codec assignment optimization. We have implemented a coarse-to-fine strategy, which significantly enhances efficiency. Notably, our framework maintains compatibility with all standard image codecs without necessitating structural modifications. Empirical results indicate that our framework establishes a new standard in LIC, advancing the Pareto frontier for performance-complexity trade-offs. It achieves a significant 70% reduction in decoding time compared to current state-of-the-art methods, without compromising reconstruction quality. Furthermore, under comparable conditions, our approach not only outperforms but significantly eclipses existing Rate-Distortion-Complexity (RDC) optimized codecs, with decoding speeds up to 30 times faster.

Infrared-visible image fusion methods aim at generating fused images with good visual quality and also facilitate the performance of high-level tasks. Indeed, existing semantic-driven methods have considered semantic information injection for downstream applications. However, none of them investigates the potential for reciprocal promotion between pixel-wise image fusion and cross-modal feature fusion perception tasks from a macroscopic task-level perspective. To address this limitation, we propose a unified network for image fusion and semantic segmentation. MAFS is a parallel structure, containing a fusion sub-network and a segmentation sub-network. On the one hand, we devise a heterogeneous feature fusion strategy to enhance semantic-aware capabilities for image fusion. On the other hand, by cascading the fusion sub-network and a segmentation backbone, segmentation-related knowledge is transferred to promote feature-level fusion-based segmentation. Within the framework, we design a novel multi-stage Transformer decoder to aggregate fine-grained multi-scale fused features efficiently. Additionally, a dynamic factor based on the max-min fairness allocation principle is introduced to generate adaptive weights of two tasks and guarantee smooth training in a multi-task manner. Extensive experiments demonstrate that our approach achieves competitive results compared with state-of-the-art methods. The code is available at https://github.com/Abraham-Einstein/MAFS/.

We introduce a novel and challenging task within the AIGC era: coding for AIGI (AI-generated images). Specifically, we propose the first AIGI dataset called PKU-AIGI-500K, which is meticulously constructed based on five major foundation models with diverse prompts. Furthermore, We conduct extensive and systematic analysis of the essential characteristics of AIGC images. We thoroughly benchmark the rate-distortion performance and runtime complexity analysis of conventional and learned image coding solutions that are openly available, revealing new insights for emerging studies in AIGI compression. The main contributions of this paper can be summarized as follows: (i) we contribute and build the first AIGI dataset, PKU-AIGI-500K, containing 105+ prompts and 528k+ images based on five generative models. Additionally, we analyze the image features that may be beneficial for image compression, such as quality, texture, color, etc., providing novel insights into available solutions and paving the way for future research. (ii) Building upon the proposed PKU-AIGI-500K, we evaluate the compression efficiency using some popular traditional codecs and learning-based codecs to form a strong benchmark. We also observe that the learning-based models trained on natural images cannot achieve competitive performance on the AIGIs without fine-tuning, emphasizing the necessity of a sufficiently large dataset to advance the research of AIGI coding. (iii) The evaluation and benchmarking are accomplished as an AIGIs’ compression project.

In this paper, a hybrid video compression framework is proposed that serves as a demonstrative showcase of deep learning-based approaches extending beyond the confines of traditional coding methodologies. The proposed hybrid framework is founded upon the Enhanced Compression Model (ECM), which is a further enhancement of the Versatile Video Coding (VVC) standard. We have augmented the latest ECM reference software with well-designed coding techniques, including block partitioning, deep learning-based loop filter, and the activation of block importance mapping (BIM) which was integrated but previously inactive within ECM, further enhancing coding performance. We evaluate the coding performance of the proposed framework with extensive experiments on the JVET dataset compared with ECM10.0 and VTM-11.0. Due to the testing environment and the coding complexity of the ECM, we did not conduct testing on Class A. The QPs are set as 22, 27, 32, 37, and 42. Compared with ECM-10.0, our method achieves 6.26%, 13.33%, and 12.33% BD-rate savings for the Y, U, and V components under random access (RA) configuration. The traditional hybrid coding framework combined with the three coding tools can further improve compression efficiency and has great potential for performance improvement.

Recent years witnessed the development in MPEG point cloud compression (PCC). However, the exploration of inter-frame coding may be impeded due to the lack of dynamic point clouds (point cloud sequences). To promote the development of PCC technology, we propose Dynamic3D , a dynamic 3D point cloud dataset with high-quality real-captured 3D persons and objects. There are several appealing properties: 1) Dynamic scenes: It contains five sequences and each sequence comprises 600 frames with temporal variation; 2) Complex content: instead of a single person or object in the existing dataset from MPEG, our established dataset contains multiple persons or both person and objects; 3) Realistic capture: the color industrial cameras and infrared cameras are used for data acquisition. This dataset provides the vast exploration space for PCC, especially the elimination of temporal redundancy. Extensive simulations are conducted on this dataset by using the reference software of MPEG G-PCC and V-PCC, i.e., (GeS-TM and TMC2), delivering observations, analysis and opportunities for the future research of PCC.

Rate control (RC) is crucial in achieving stable and smooth bitrate variation in video compression and transmission. Existing RC methods for neural video compression (NVC) have made strong assumptions on solving bit allocation parameters using a pre-defined model, leading to high bit-rate errors (BRE). In response, the study introduces a simple yet effective one-pass RC strategy tailored for NVC frameworks in a plug-in fashion. This strategy consists of two key components: the NVC rate-adaptive model and the associated RC approach. The former model constructs the basis of the latter approach. The proposed RC approach employs a progressive online updating technique for parameter estimation to achieve a lower BRE and maintain the original quality structure of frameworks. Experimental results demonstrate that our approach achieves an impressive RC performance on standard test sequences, outperforming the conventional optimal R-λ RC model with lower BRE and better rate-distortion (R-D) performances. Furthermore, we extend our method to three representative NVC frameworks, consistently showcasing its effectiveness in achieving lower BRE with only moderated R-D performance degradation.

The widespread adoption of bio-inspired cameras has catalyzed the development of spike-based intelligent applications. Despite its innovative imaging principle allows for functionality in extreme scenarios, the intricate nature of spike signals poses processing challenges to achieve desired performance. Traditional methods struggles to deliver visual perception and temporal prediction simultaneously, and they lack the flexibility needed for diverse intelligent applications. To address this problem, we analyze the spatio-temporal correlations between spike information at different temporal scales. A novel spike processing method is introduced for compact spike representations that utilizes intra-scale correlation for higher predictive accuracy. Additionally, we propose a multi-scale spatio-temporal aggregation unit (MSTAU) that further leverages inter-scale correlation to achieve efficient perception and precise prediction. Experimental results show noticeable improvements in scene reconstruction and object classification, with increases of 3.49dB in scene reconstruction quality and 2.20% in accuracy, respectively. Besides, the proposed method accommodate different visual applications via switching analysis models, offering a novel perspective for spike processing.

Deep video compression has attracted increasing attention in recent years due to its end-to-end optimization ability. However, most existing neural video compression (NVC) models focus on incorporating sophisticated motion or residual coding networks for successive frames leveraging spatial-temporal redundancy removal, neglecting the efficient motion representation and essential structure of scaled prediction for motion dynamics. To resolve this problem, this paper proposed a novel model, named scaled hierarchical bi-directional prediction structure, which effectively captures temporal correlation among frames considering the quality variation when managing the reference frames. This paper first introduces parameter-shared motion codecs and efficient information fusion strategies to obtain predictive features more precisely. Subsequently, scaled motions from temporal contexts are learned as bi-directional prior for motion representation. Additionally, the concept of trustworthy motion modeling is proposed to represent the effectiveness of reference information, measuring the reliability of predictive accuracy in complex motions, camera rotations and occlusions. Extensive experimental results demonstrate that our approach offers significant advantages over state-of-the-art bi-directional NVC models in coding efficiency. The proposed method has been adopted as the latest reference model by Moving Picture, Audio, and Data Coding by Artificial Intelligence (MPAI) end-to-end video coding (EEV) standard. The code is available at: https://github.com/yefeng00/DVC_with_Scaled_Hierarchical_Bi_directional_Motion_Model.

End-to-end image compression has achieved satisfactory results in recent studies. However, existing methods suffer from high complexity of complicated neural network computation and cannot be directly deployed on mobile devices due to the limitations of computing ability and storage. Therefore, considering the resource and computing ability constrains of the mobile devices, we make a trade-off in this paper between rate-distortion (R-D) performance, inference time, and model complexity. Then we design a novel lightweight perceptual image compression framework to alleviate the storage and complexity burden of mobile devices. Moreover, we design a hardware-friendly deployment scheme to apply the proposed compression framework on high-end mobile devices, which can achieve efficient image compression. Based on the above structures, we propose the first mobile system that achieves image compression on mobile devices. The supplementary material of our system demo is on https://sigport.org/documents/extreme-low-bitrate-Image-compression-system-mobile-deployment.

Recent advancements in telecommunication technologies have exceeded the requirements of numerous video-based Real-Time Communication (RTC) applications. Mean-while, in light of the growing demand for immersive 3D visual experience, researchers are currently focusing on developing next-generation telepresence systems. This paper presents a novel immersive conferencing system that offers a smooth, high-fidelity, and life-size autostereoscopic display of remote user portraits. To generate the binocular stereo vision in real time, an adaptive low-complexity view synthesis method based on an accelerated Real-time Intermediate Flow Estimation (RIFE) model is employed, which performs direct cross-view generation based on decoded multi-view videos and tracked eye positions of the watching user. Thanks to eliminating the complex 3D modeling procedure that relies on depth images, the proposed system requires significantly fewer computational resources and lower video transmission bandwidth compared to existing immersive conferencing systems. Therefore, the proposed system is low-cost and flexible when accommodating diverse conferencing environments.

As a novel bio-inspired imaging device, the spike camera shows remarkable potential in capturing ultra-high-speed motion scenes by simulating the mechanism of the retinal fovea. It achieves a temporal resolution of tens of thousands of Hz through spike emission. However, this capability presents significant challenges in terms of large-scale data storage and transmission, along with stringent fidelity requirements for spike data, thus posing a formidable obstacle to the lossless compression of continuous spike streams. In this paper, we introduce an effective image representation method for spikes, along with an intensity remapping technique to mitigate noise effects in spike streams. Building on this, we propose a learned lossless spike data compression model. To our knowledge, it is the first learning-based model for lossless spike stream compression. Experimental results demonstrate that our method can realize state-of-the-art performance for spike data lossless compression.

In the fourth generation Audio Video coding Standard (AVS4), the Inter Prediction Filter (INTERPF) reduces discontinuities between prediction and adjacent reconstructed pixels in inter prediction. The paper proposes a low complexity learning-based inter prediction (LLIP) method to replace the traditional INTERPF. LLIP enhances the filtering process by leveraging a lightweight neural network model, where parameters can be exported for efficient inference. Specifically, we extract pixels and coordinates utilized by the traditional INTERPF to form the training dataset. Subsequently, we export the weights and biases of the trained neural network model and implement the inference process without any third-party dependency, enabling seamless integration into video codec without relying on Libtorch, thus achieving faster inference speed. Ultimately, we replace the traditional handcraft filtering parameters in INTERPF with the learned optimal filtering parameters. This practical solution makes the combination of deep learning encoding tools with traditional video encoding schemes more efficient. Experimental results show that our approach achieves 0.01%, 0.31%, and 0.25% coding gain for the Y, U, and V components under the random access (RA) configuration on average.

Previous research demonstrated that filtering Macro-Pixels (MPs) in a decoded Light Field Image (LFI) sequence can effectively enhances the quality of the corresponding Sub-Aperture Images (SAIs). In this paper, we propose a deep-learning-based quality enhancement model following the MP-wise processing approach tailored to LFIs encoded by the Versatile Video Coding (VVC) standard. The proposed novel Res2Net Quality Enhancement Convolutional Neural Network (R2NQE-CNN) architecture is both lightweight and powerful, in which the Res2Net modules are employed to perform LFI filtering for the first time, and are implemented with a novel improved 3D-feature-processing structure. The proposed method incorporates only 205K model parameters and achieves significant Y-BD-rate reductions over VVC of up to 32%, representing a relative improvement of up to 33% compared to the state-of-the-art method, which has more than three times the number of parameters of our proposed model.

Video Coding for Machines (VCM) aims to compress visual signals for machine analysis. However, existing methods only consider a few machines, neglecting the majority. Moreover, the machine's perceptual characteristics are not leveraged effectively, resulting in suboptimal compression efficiency. To overcome these limitations, this paper introduces Satisfied Machine Ratio (SMR), a metric that statistically evaluates the perceptual quality of compressed images and videos for machines by aggregating satisfaction scores from them. Each score is derived from machine perceptual differences between original and compressed images. Targeting image classification and object detection tasks, we build two representative machine libraries for SMR annotation and create a large-scale SMR dataset to facilitate SMR studies. We then propose an SMR prediction model based on the correlation between deep feature differences and SMR. Furthermore, we introduce an auxiliary task to increase the prediction accuracy by predicting the SMR difference between two images in different quality. Extensive experiments demonstrate that SMR models significantly improve compression performance for machines and exhibit robust generalizability on unseen machines, codecs, datasets, and frame types.

Cross-modal compression (CMC) aims to compress highly redundant visual data into compact, common, and human-comprehensible domains, such as text, to preserve semantic fidelity. However, CMC is limited by a constant level of semantic fidelity and constrained semantic fidelity due to a single compression domain (plain text). To address these issues, we propose a new approach called Multiple-domains rate-distortion optimized CMC (M-CMC). Specifically, our method divides the image into two complementary representations: (1) a structure representation with an edge map, and (2) a texture representation with dense captions, which include numerous region-caption pairs instead of plain text. In this way, we expand the single domain to multiple domains, namely, edge maps, regions, and text. To achieve diverse levels of semantic fidelity, we suggest a rate-distortion reward function, where the distortion measures the semantic fidelity of reconstructed images and the rate measures the information content of the text. We also propose Multiple-stage Self-Critical Sequence Training (MSCST) to optimize the reward function. Extensive experimental results demonstrate that the proposed method achieves diverse levels of semantic translation more effectively than other CMC-based methods, achieves higher semantic compression performance compared to traditional block-based and learning-based image compression frameworks with 97,000-500 times compression ratio, and provides a simple yet effective way for image editing.

In recent years, Learned Image Compression (LIC) has undergone rapid evolution. However, it is worthy noting that most prevalent LIC methodologies still rely on uniform Scalar Quantization (SQ) for latent features. This overlooks the untapped potential of contextual information, which could be leveraged to significantly reduce statistical redundancies. Prior researches have explored Vector Quantization (VQ)’s adaptability to diverse data distributions, yet it introduces significant computational complexity into LIC, hindering its practical implementation. Consequently, in this work, we propose the Contextual Sequential Quantization (CSQ) method, which progressively discretizes the latent features of LIC by harnessing content contextual information and image textural priors. Our proposed CSQ signifies progress in LIC by blending the computational efficiency of SQ with a substantial approach towards the adaptability of VQ. We further propose the Center Compensation Module (CCM) based on the proposed CSQ. This module strategically determines adaptive quantization centers, leading to a direct enhancement of reconstruction quality without compromising the bit-rate. Moreover, it is worth noticing that existing LIC approaches face challenges in leveraging hyper side information to effectively enhance transformations, which is attributed to the entanglement of the hyperprior generation module with the main transformations. Consequently, we propose to decouple the hyperprior module from main transformations, and design the Hyperprior-Assisted Transformation (HAT) unit to feed hyperprior back into main transformations. This further improves the coding performance. By integrating all together the proposed CSQ, CCM, and HAT, our proposed Non-uniform quantization-based LIC (NLIC) method attains state-of-the-art rate-distortion (R-D) performance among existing LIC methodologies.

The rapid advancement of artificial intelligence (AI) technology has led to the prioritization of standardizing the processing, coding, and transmission of video using neural networks. To address this priority area, the Moving Picture, Audio, and Data Coding by Artificial Intelligence (MPAI) group is developing a suite of standards called MPAI-EEV for "end-to-end optimized neural video coding." The aim of this AI-based video standard project is to compress the number of bits required to represent high-fidelity video data by utilizing data-trained neural coding technologies. This approach is not constrained by how data coding has traditionally been applied in the context of a hybrid framework. This paper presents an overview of recent and ongoing standardization efforts in this area and highlights the key technologies and design philosophy of EEV. It also provides a comparison and report on some primary efforts such as the coding efficiency of the reference model. Additionally, it discusses emerging activities such as learned Unmanned-Aerial-Vehicles (UAVs) video coding which are currently planned, under development, or in the exploration phase. With a focus on UAV video signals, this paper addresses the current status of these preliminary efforts. It also indicates development timelines, summarizes the main technical details, and provides pointers to further points of reference. The exploration experiment shows that the EEV model performs better than the state-of-the-art video coding standard H.266/VVC in terms of perceptual evaluation metric.

This paper proposes a unified and efficient entropy coding method for learned image compression (LIC) from the perspective of traditional signal processing. First, the consistency of structures and optimization objectives are used to interpret the existing split-coded-then-merge entropy coding strategies in LIC as a particular filter banks framework, with feature separation and feature aggregation representing the analysis filter bank and synthesis filter bank, respectively. Thus, we borrow the design from the multirate filter banks and proposed Multirate Progressive Entropy Model (MPEM) to enhance the rate-distortion performance and decoding speed. In particular, we create an analysis filter bank that divides compact features into a few nonuniform subsets based on various spatial and channel sampling rates. Then multi-scale detail and mean coefficients within the current subset are used as prior representations to help generate the prediction parameters of the next subset, and the carefully designed synthetic filter bank performs a near-perfect reconstruction of the features. In addition, we propose a Multi-level Edge Attention Moudal (MEAM) to increase the edge and texture information’s contribution and reduce the high-frequency information loss brought on by MPEM’s inherent multi-rate spatial sampling, which leverages the edge operator and structural reparameterization principles. The results of the experiments show that, in comparison to the effective LIC methods and traditional code, the proposed MPEM can decode data at a cutting-edge speed while also offering comparable rate-distortion performance.

Traditional image/video compression compresses the highly redundant visual data while preserving signal fidelity. Recently, cross modal compression (CMC) is proposed to compress the data into a compact, human-comprehensible domain (such as text) with an ultra-high compression ratio while preserving semantic fidelity for machine analysis and semantic monitoring. CMC is with a constant rate because the CMC encoder can only represent the data with a fixed grain. But in practice, variable rate is necessary due to the complicated and dynamic transmission conditions, the different storage mediums, and the diverse levels of application requirements. To deal with this problem, in this paper, we propose variable rate cross modal compression (VR-CMC), where we introduce variable rate prompt to represent the data with different grains. Variable rate prompt is composed of three strategies. Specifically, 1) target length prompt (TLP) introduces the target length into the language prompt to guide the generation of the text representation; 2) decaying EOS probability (DEP) exponentially decays the probability of the EOS token with regard to the decoding step and target length, where the EOS (end-of-sequence) token is a special token indicating the end of the text; 3) text augmentation (TA) enriches the training data and makes the text representation length more balanced when training. Experimental results show that our proposed VR-CMC can effectively control the rate in the CMC framework and achieve state-of-the-art performance on MSCOCO and IM2P datasets.

Recently, cross modal compression (CMC) is proposed to compress highly redundant visual data into a compact, common, human-comprehensible domain (such as text) to preserve semantic fidelity for semantic-related applications. However, CMC only achieves a certain level of semantic fidelity at a constant rate, and the model aims to optimize the probability of the ground truth text but not directly semantic fidelity. To tackle the problems, we propose a novel scheme named rate-distortion optimized CMC (RDO-CMC). Specifically, we model the text generation process as a Markov decision process and propose rate-distortion reward which is used in reinforcement learning to optimize text generation. In rate-distortion reward, the distortion measures both the semantic fidelity and naturalness of the encoded text. The rate for the text is estimated by the sum of the amount of information of all the tokens in the text since the amount of information of each token is a lower bound of coding bits. Experimentally, RDO-CMC effectively controls the rate in the CMC framework and achieves competitive performance on MSCOCO dataset.

In this paper, we propose to build a novel benchmark and neural video coding task named learning based Unmanned Aerial Vehicle (UAV) video coding. We collect the UAV videos with different content variations, including in-door and out-door scenes, object-scale variations and viewpoint distance, different climate condition etc. Then we encode those properly-selected videos using popular end-to-end optimized video codecs and conventional hybrid codecs, to form a comprehensive benchmark for learned drone video compression. We also provide a detailed analysis and envision the challenge of such task for future research. The main contributions of this paper are three folds. First, we construct a comprehensive benchmark for the task of drone video compression which consists of the rate-distortion (R-D) behavior of both hybrid and learned video codecs. To our knowledge, it is the first attempt in end-to-end optimized solution to compress drone videos. Second, we provide the review and analysis of the learned drone video compression schemes and further discuss the challenges of encoding UAV videos. Third, this benchmark and related research is accomplished as a milestone MPAI End-to-end Video (EEV) coding project. The proposed benchmark has constructed a solid baseline for compressing UAV videos and facilitates the future research works for related task.

Learned Image Compression (LIC) has surpassed traditional image compression in terms of Rate-Distortion (R-D) performance with strong and powerful non-linear transformations. However, most existing LIC approaches discretize the latent representation into integers. Such method remains the statistical redundancy and conditional information of image content, which could be further leveraged to systematically reduce bit rate in the lossy compression regime. In this work, we formulate the problem of Non-Uniform Conditional Quantization (NUCQ) in LIC and accordingly propose to learn a content-adaptive NUCQ scheme conditioned on image textural prior information. In our method, a parameterized quantile model is learned from the hyperprior information, thus more content adaptive quantization is realized in the latent domain. Experiments demonstrate that our proposed NUCQ improves the R-D performance of LIC by 4.83% and 6.03% on the Bjøntegaard-delta bit rate (BD-rate) metric concerning the MSE and MS-SSIM metric, respectively.

The increasing popularity of video conferencing and live streaming raises the growing demand for encoding human-oriented videos at ultra-low bit rates. Recently, several ultra-low bitrate video codecs have proposed using inter-frame keypoints or landmarks to derive motion representations, which are then used to warp decoded frames in a generative manner. Despite its success, compression of the motion representation has been less investigated in the literature. In this work, we propose a novel principal component analysis (PCA)-based decomposing method to fully exploit the compression potential of motion representations. In particular, we decompose the derived motion affine matrices into three parts and apply quantization and entropy estimation to each part in a different way depending on its significance. Using such compressed-friendly motion representations allows for preserving most of the motion information and achieving lower coding costs. Extensive qualitatively and quantitatively experimental results on the human video datasets demonstrate the superiority of the proposed paradigm over existing video codecs under extreme compression ratios.

Audio Video Coding Standard (AVS) Intelligent Coding Group mainly studies video coding tools based on neural network technology and its potential benefit for next generation video coding. Extensive efforts have been dedicated to the research on neural network (NN) based coding tools. In this paper, we present a novel NN based video coding framework by leveraging the supervised trained NN models for multiple modules in the hybrid coding framework, from the predictive coding to the in-loop filtering. Specifically, NN based intra prediction models the non-linear mapping from contextual pixels to the predictions. The inter prediction efficiency is enhanced by introducing a virtual reference frame (VRF) network. The convolutional neural network based loop filtering (CNNLF) with discriminative model selection exploits the texture adaptivity. The experimental results show that the CNNLF, NN Intra, and VRF models can bring 8.60%, 1.02%, and 2.26% luma BD-rate reduction under random access (RA) configuration compared with AVS reference software HPM13.0. Additional experiments with the combined three NN coding tools reveal that around 13% YUV BD-rate reduction could be obtained. The proposed framework opens novel sights for next generation video coding from the intelligent coding perspective.

This paper presents an innovative pipeline design for the quantization of neural image codecs, ensuring robust and consistent deployment across various platform inferences. Our approach addresses a significant challenge prevalent in neural image compression networks: the lack of quantization methods for entropy probability hyper-priors, such as Gaussian mixture models (GMMs), in the existing literature. Our method serves as a solution to prevent decoding errors stemming from float-point discrepancies between the encoder and decoder, especially when executed on different platforms. By implementing a quantization scheme based on look-up tables (LUTs), we achieve precise quantization of GMM probability models with minimal performance degradation, approaching near losslessness. Moreover, our contribution offers a comprehensive quantization pipeline that facilitates dependable data transmission and harmonious synchronization of arithmetic coding algorithms between the encoder and decoder components. Empirical evidence highlights the effectiveness of our method, demonstrating seamless codec matching on diverse platforms and substantiating its capacity to enhance the integrity and security of compressed data. This work enhances the practical applicability of neural compression techniques across real-world scenarios.

Most of the existing deep learning-based video compression frameworks rely on motion estimation and compensation. However, the artifacts of the warped frames after motion compensation, which propagate the errors to the next frame, limit the video coding performance. In this work, we propose enhanced motion compensation for reduced error propagation in deep video compression. More specifically, we incorporate the designed convolutional neural network into Open DVC as the motion compensation enhancement network to remove noise in the predicted frame. With the enhanced frame, we jointly optimize the whole framework with a single loss function by considering the trade-off between bit cost and frame quality. Experiments show that the proposed enhanced motion compensation model reduces error propagation within a group of frames. Compared with Open DVC, our model can achieve 8.94% bit savings on average for standard test videos in terms of PSNR. Regarding MS-SSIM, our model outperforms Open DVC with 5.67% bit rate savings.

Inter prediction is the critical component in hybrid coding framework to deal with the temporal redundancy. Most of the neural video coding methods typically follow the motion compensation based inter coding scheme, establishing motion vector (MV) as the central role. In this paper, we innovatively propose an efficient motion modeling approach by inherently decomposing it into two components, the intrinsic motion and the compensatory motion. The intrinsic motion originates from the implicit spatiotemporal context hidden in the historical sequence, which can be intuitively captured free of bits. On the top of it, the compensatory motion acts a role of structural refinement and texture enhancement as a form of side information. In particular, the inter prediction is performed in the feature space as a manner of progressive temporal transition, conditioned on the decomposed motion. By the motion decomposition paradigm, we innovatively answer the question of motion representation, compensation and coding in the learned video compression framework. With the temporal prediction, the remaining pixel residue is signaled to obtain the reconstruction. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art coding performance on par with other end-to-end coding methods, and outperforms versatile video coding (VVC) under low-delay P (LDP) configuration in terms of MS-SSIM metric.

Learned image compression methods have achieved satisfactory results in recent years. However, existing methods are typically designed for RGB format, which are not suitable for YUV420 format due to the variance of different formats. In this paper, we propose an information-guided compression framework using cross-component attention mechanism, which can achieve efficient image compression in YUV420 format. Specifically, we design a dual-branch advanced information-preserving module (AIPM) based on the information-guided unit (IGU) and attention mechanism. On the one hand, the dual-branch architecture can prevent changes in original data distribution and avoid information disturbance between different components. The feature attention block (FAB) can preserve the important information. On the other hand, IGU can efficiently utilize the correlations between Y and UV components, which can further preserve the information of UV by the guidance of Y. Furthermore, we design an adaptive cross-channel enhancement module (ACEM) to reconstruct the details by utilizing the relations from different components, which makes use of the reconstructed Y as the textural and structural guidance for UV components. Extensive experiments show that the proposed framework can achieve the state-of-the-art performance in image compression for YUV420 format. More importantly, the proposed framework outperforms Versatile Video Coding (VVC) with 8.37% BD-rate reduction on common test conditions (CTC) sequences on average. In addition, we propose a quantization scheme for context model without model retraining, which can overcome the cross-platform decoding error caused by the floating-point operations in context model and provide a reference approach for the application of neural codec on different platforms.

Low bit-width quantization can effectively reduce the storage and computational costs of deep neural networks. Existing quantization methods are commonly designed for single model compression. For multi-model compression scenarios, multiple models for the same task or similar tasks need to be compressed simultaneously in multimedia tasks, such as compressing image super-resolution models for different scales and transferring of different models in multimedia. However, single model quantization methods do not consider the correlations among the weights of different models, which limits the further compression for the above multi-model compression scenarios. To sufficiently excavate the potential of compression on multi-model, we propose a novel quantization scheme for multi-model compression, namely differential weight quantization (DWQ), which focuses on the weights increment between the target model and the reference model. Specifically, DWQ is achieved by increment computation, increment quantization and fine-tuning, which utilizes the reference model to guide the subsequent quantization on the target model. Due to the correlations between the weights of different models, the distribution of weights increment is more centralized compared with original weights, which can achieve a higher compression ratio by lower bit representation on weights increment. Moreover, the progressive training method is proposed to accelerate the convergence and reduce quantization loss on the DWQ framework. Extensive experiments validate the effectiveness of DWQ based on weight-sharing and parameterized clipping activation (PACT) quantization technologies on multiple tasks. The proposed framework can achieve 2× compression improvement and reduce 30% computational complexity with comparable performance in the popular multimedia tasks.

In this paper, we propose a novel low-complexity in-loop filtering approach named textural and directional information based offset (TDIO) for the video coding standard AVS3. Different from conventional offset-based filtering methods which partially use contextual samples, the key contribution of TDIO is that it fully utilizes the textural and edge directional features of each sample to comprehensively determine which type of texture characteristics each sample belongs to. The corresponding offsets are generated and signaled to decoder such that sample-level distortion is reduced. Specifically, the multi-directionality and sample-intensity pattern based classifiers are first proposed to extract the directional and textural features, respectively. The classification results are obtained by incorporating these features, and the optimal offset values for each class are derived based on rate-distortion optimization. Since sample-level offset signalling may cause heavy burden to the overhead of TDIO, we subsequently propose a filtering offset sharing mechanism based on historical information between available temporal-adjacent compressed frames. In addition, an iteration-based filter adaptation method is designed to improve the local adaptivity of TDIO for better compression efficiency. Experimental results show that the proposed TDIO achieves 0.64%, 1.29%, 1.86%, and 2.20% bit rate savings for all intra, random access, low delay B, and low delay P configurations, respectively. Moreover, TDIO is helpful to improve subjective quality by leveraging the fine-grained local texture characteristics. It can be observed that the blurring and ringing artifacts could be significantly suppressed by using the proposed method, yielding higher subjective quality.

End-to-end optimized neural image compression (NIC) has obtained superior lossy compression performance recently. In this paper, we consider the problem of rate-distortion (R-D) characteristic analysis and modeling for NIC. We make efforts to formulate the essential mathematical functions to describe the R-D behavior of NIC using deep networks. Thus arbitrary bit-rate points could be elegantly realized by leveraging such model via a single trained network. We propose a plugin-in module to learn the relationship between the target bit-rate and the binary representation for the latent variable of auto-encoder. The proposed scheme resolves the problem of training distinct models to reach different points in the R-D space. Furthermore, we model the rate and distortion characteristic of NIC as a function of the coding parameter$\lambda$respectively. Our experiments show our proposed method is easy to adopt and realizes state-of-the-art continuous bit-rate coding performance, which implies that our approach would benefit the practical deployment of NIC.

Recently many efforts have been devoted to learning non-linear predictions from neighboring samples with deep neural networks. However, existing methods mainly generate predictions with local reference samples, regardless of nonlocal self-similarity. In this paper, we aim to incorporate local and nonlocal correlations for intra prediction and propose a two-stage coarse-to-fine network (CTFN), which is integrated into VVC codec as an optional intra prediction mode. The prediction process of CTFN is decomposed into two stages. In the first stage, we train a set of networks to generate a coarse result with local reference samples. In the second stage, we extract sufficient features from nonlocal region using the coarse result as priors and transform the features into a fine prediction result. In particular, a patch-wise attention layer (PAL) is designed in the second stage that can fully explore nonlocal correlations in feature domain and assign weights to each nonlocal feature adaptively, as shown in Fig. 1. As such, the proposed CTFN can not only learn a non-linear mapping from local context, but also explicitly borrow similar features from nonlocal region in a weighted form. Different from image inpainting tasks, the patch synthesis problem is converted to patch matching problem with the CTFN, yielding more reliable predictions. More-over, we construct a classified dataset based on Pearson Correlation Coefficient for network training to better handle contents that are highly correlated. Experiments on VTM-11.0 show that the proposed network achieves 1.77% ED-rate reductions under all intra configuration, which outperforms the state-of-the-art methods.

Intra prediction acts a significant role in removing the spatial redundancy in the hybrid coding framework. Versatile Video Coding (VVC) employs a set of angular intra modes to generate directional contents based on the linear projection hypothesis. However, the linear based predictor is not expressive enough for generating high-fidelity patterns with non-linear structure. To compensate for that, we propose a high-order intra prediction (HOIP) for future video coding in this paper. In particular, the HOIP is modeled by a quadratic extrapolation function. To be compatible with the present intra prediction mechanism, the quadratic function can be formulated by two angular intra modes. To reduce the encoding complexity, we further propose a search pruning strategy to find the most appropriate pair-wise modes, and it can be flexibly extended for higher coding performance. The extensive experimental results demonstrate the effectiveness of the proposed method. Up to 0.6% BD-rate saving is obtained with the moderate complexity increment.

In-loop filter has been comprehensively explored during the development of video coding standards to suppress compression artifacts. However, the existing in-loop filters in Versatile Video Coding (VVC) mainly take advantage of the image local similarity. Although some non-local based in-loop filters can make up for this short-coming, the unsupervised parameter selection scheme, which is widely used by non-local filters, limits the content adaptability. Given this, we propose a parametric non-local in-loop filter (PNLF) that fully considers the non-local characteristics and trains the filter coefficients based on the video content. In the filtering process, the reference samples based on the non-local similarity are first derived for each to-be-filtered sample. Then to-be-filtered samples are grouped into specific classes based on multiple features. For each class, filter coefficients are online trained in the encoder and transmitted to the decoder. Finally, the filtering process is conducted using the online-selected coefficients. Simulation results reveal that the proposed approach achieves 0.70%, 1.43%, and 2.09% bit-rate savings on average compared to VTM-11.0 under All Intra (AI), Random Access (RA), and Low-Delay B (LDB) configurations, respectively. The sequences used in the experiment include Class AI, A2, B, C, D, E, F, and SCC. Compared to the non-local structure-based filter (NLSF) [1], our proposed PNLF with fast block matching scheme [2] applied on B-frames and P-frames can achieve better performance gain with lower software and hardware complexity under RA and LDB configurations.

Flexible coding unit partitioning such as quad-tree nested binary-tree and ternary-tree adopted by the emerging enhanced compression model (ECM) brings promising coding performance improvement. Meanwhile, the computational complexity increases dramatically, which may block the exploration and validation of new coding tools. This paper investigates a partition mode early pruning scheme via a fully connected network to reduce the encoding complexity for the ECM. In particular, we carefully select features and devise the fully connected network, which could seamlessly cooperate with the encoder, revealing promising learning and inference capability. Experimental results demonstrate that the proposed method achieves 15%~50% encoding time savings with moderate bit-rate increasing on the ECM, and the extra complexity regarding the fully connected network and feature extraction is negligible.

Image compression approaches based on deep learning have achieved remarkable success. Existing studies mainly focus on human vision and machine analysis tasks taking reconstructed images as input. However, those methods need images to be decoded before performing downstream visual tasks, which motivates us to explore how to directly conduct visual analysis using the compressed data without decoding. The overview of our proposed model is shown as Fig. 1(a). Specifically, a task-agnostic learning-based compression model is proposed, which effectively supports various compressed domain-based analytical tasks meanwhile reserves outstanding re-constructed perceptual quality compared with traditional and learning-based codecs. To obtain the extremely compacted data representation with essential semantic infor-mation, we take the help of the generative model on decoder part. Then, we propose a multi-task learning model which can directly obtain semantic information from the compressed visual data. The pipeline of the proposed model is detailedly illus-trated in Fig. 1(b). In addition, joint optimization strategy is adopted to achieve the best balance point among compression efficiency, reconstructed image quality, and the downstream visual tasks' performance. Experimental results verify that our proposed compressed domain-based multi-task analysis model outperforms the reconstructed image-based method on transmission efficiency, saving more than ten times of bit-rate consumption while preserving comparable visual analysis precision (i.e., classification and segmentation tasks) when compared with RGB image input models, which is evaluated on the CelebA-HO dataset.

Image compression is a key problem in this age of information explosion. With the help of machine learning, recent studies have shown that learning-based image compression methods tend to surpass traditional codecs. Image compression can be split into three steps: transform, quantization, and entropy estimation. However, the transform step in traditional codecs lacks flexibility because of the strict mathematical premise while the transform in most learning-based codecs neglects its intrinsic interpretation. After observing compression degradation degree varies on different frequency bands as illustrated as Fig. 1(a), we propose an end-to-end compression model from the frequency perspective with a frequency-pyramid transform and a frequency-aware fusion module. The right of the Fig. 1(a) displays each frequency layer's component of the proposed model from low to high-frequency splits. Intuitively, we can infer that the low-frequency part contains the global structure while the high-frequency part gets finer details, satisfying the feature of human visual system (HVS). The proposed model are detailedly shown in Fig. 1(b) that independent probability estimation models are set for each frequency split. Extensive experiments are conducted to demonstrate that our model outperforms all traditional codecs (e.g., JPEG, JPEG2000, HEVC, and VVC) on MS-SSIM metric on both Kodak and CLIC2020 professional test datasets. Taking BPG-4:4:4 as the anchor, our proposed model achieves 11.6% BD-rate reduction under PSNR measurement, which is evaluated on the Kodak dataset.

Deep learning-based end-to-end image compression has achieved significant compression performance in recent years. However, current learning-based image compression methods are designed considering the characteristic of RGB color space, which is not suitable for image compression in YUV 420 color space because of the variance between color formats. To achieve efficient image compression in YUV 420 color space, we propose an information-preserving compression framework using the attention mechanism. Specifically, we design an information-preserving module (IPM), where we utilize the dual-branch architecture to prevent changes in data distribution and propose the feature attention block (FAB) to preserve information. Furthermore, a cross-channel progressive enhancement (CPE) network is designed by taking advantage of the relations among different channels. Ex-perimental results show that the proposed framework outper-forms state-of-the-art compression standard Versatile Video Coding (VVC) with 2.52% BD-rate reduction on common test conditions (CTC) sequences on average.

Recent years have witnessed the great promise of deep neural video compression codecs. However, there are still unprecedented challenges ahead when the videos are expected to be encoded with extremely low bitrate. Motivated by recent attempts of layered conceptual image compression, we make the first attempt to leverage the disentangled visual representations for extreme human body video compression. More specifically, to capture the main structure, we adopt the inferred human pose keypoints as the structure code of each frame, thereby deriving the motion information from structure codes of adjacent frames for further compression. To better exploit the texture redundancy, all frames share the same texture codes by incorporating the proposed texture contrastive learning to ensure texture consistency within a video. Two branches are consequently transmitted in a separable manner, and the generator synthesizes the reconstructed video with the combination of all decoded representations at the decoder side. Both qualitative and quantitative experimental results demonstrate that the proposed scheme can produce perceptually pleasing reconstruction results in ultra-low bitrates far below that can be reached by other video codecs.

Hardware-oriented neural video coding (NVC) has been an emerging research topic recently. In this work, we present a P-frame based NVC system by leveraging the parallel neural inference capability of resource-constrained embed platforms, such as field programmable gate array (FPGA). The proposed system, namely FPX-NVC, consists of multiple resolution video capture, reconfigurable and integerized neural video codec and display. Our neural codec enjoys the learned unidirectional inter-predictive coding by jointly optimizing the implicit motion representation (IMR), feature domain motion compensation, and the corresponding prediction error compression, realizing end-to-end on-board encoding and decoding. Our system supports end-to-end video data operation flow from capturing to neural coding and display. To our knowledge, the FPX-NVC is the first inter-predictive coding (P-frame) based hardware NVC system.

The recent trend in neural image compression (NIC) research could be generally grounded into two categories: analysis-synthesis transform network improvements and entropy estimation optimization. They promote the compression efficiency of NIC by leveraging more expressive network structures and advanced entropy models respectively. From a different but more systematic viewpoint, we extend the horizon of NIC from software- to hardware-based lossy compression using more resource-constrained platforms, such as field programmable gate array (FPGA) or deep-learning processor unit (DPU). In this paper, we propose a novel hardware-oriented NIC system for real-time edge-computing video services. We for the first time present FPX-NIC, an FPGA-accelerated NIC framework designed for hardware encoding, which consists of a novel NIC scheme and an energy-efficient neural network (NN) deployment method. The former contribution is a block-based adaptive NIC approach based on local content characteristics. Essential side-information is also signalled to realize adaptive patch representation. The critical advantage of our latter contribution lies in the network-reconfigurable framework plus fixed-precision weights quantization method that takes advantage of quantization-aware post training procedure to compensate the performance degradation caused by quantization error. Therefore it is able to improve both processing speed and energy efficiency. We finally establish an intelligent video coding system using the proposed scheme, enabling visual capturing, neural encoding, decoding, and display, realizing 4K ultra-high-definition (UHD) all intra neural video coding on edge-computing devices.

Dynamic textures are time-varying motion patterns that exhibit certain temporal stationarity. The motion of such patterns is usually anisotropic, which is a great challenge for video coding. In this paper, we firstly analyze the predictive and temporal characteristics of dynamic textures and, based on the analysis, a pixel-level segmentation-synthesis framework for dynamic texture compression is proposed to improve predictive-coding efficiency. This framework consists of three sub-modules: dynamic texture segmentation, dynamic texture synthesis and fusion process. A deep-learning-based segmentation method is introduced to obtain a pixel-level mask. Subsequently, optical flow fields are used to generate flowlines and help analyze the motion periodicity of a given sequence. The flowlines and period values are utilized as inputs for flow-based dynamic texture synthesis. The segmented dynamic texture area is then reconstructed using the synthesized results at the decoder side at pixel-level, such that the overhead for transmitting these contents could be saved. The segmented mask is transmitted through an All-Zero-Run-Length coding algorithm and an effective fusion module is further proposed to reduce edge artifacts that occur at the boundary of dynamic texture areas. Coding blocks have been classified into three types and a weighting reconstruction scheme is performed. The proposed framework has been integrated into VTM-10.0. Experimental results on a collected dynamic texture test set demonstrate that 15.1% and 16.1% MOS BD-rate savings can be achieved on average for LDB and LDP configurations, respectively.

Flexible block partition and newly adopted intra coding tools bring significant performance improvement for next generation video coding, and meanwhile introduce non-negligible encoding complexity increment. This paper presents a scalable intra coding optimization (SICO) scheme for the third generation of audio video coding standard (AVS3). The complexity distribution and inheriting relationship among different partitioning, as well as intra prediction modes, are systematically analyzed. Subsequently, low-complexity algorithms are proposed, which could early exclude unlikely coding modes safely. In particular, a data-driven binary classifier is elegantly trained for the determination of coding unit partitioning. Moreover, the preliminary coding information can be exhaustively utilized for the mode decision in a low-cost manner. The proposed method provides scalable intra coding optimizing solutions, which is eligible to cater to various application scenarios. Experimental results show that the proposed method could achieve a wide range of encoding complexity reduction from 18% to 76% with moderate compression performance loss. One implementation of the proposed method has been adopted to the AVS3 reference software.

Existing compression methods typically focus on the removal of signal-level redundancies, while the potential and versatility of decomposing visual data into compact conceptual components still lack further study. To this end, we propose a novel conceptual compression framework that encodes visual data into compact structure and texture representations, then decodes in a deep synthesis fashion, aiming to achieve better visual reconstruction quality, flexible content manipulation, and potential support for various vision tasks. In particular, we propose to compress images by a dual-layered model consisting of two complementary visual features: 1) structure layer represented by structural maps and 2) texture layer characterized by low-dimensional deep representations. At the encoder side, the structural maps and texture representations are individually extracted and compressed, generating the compact, interpretable, inter-operable bitstreams. During the decoding stage, a hierarchical fusion GAN (HF-GAN) is proposed to learn the synthesis paradigm where the textures are rendered into the decoded structural maps, leading to high-quality reconstruction with remarkable visual realism. Extensive experiments on diverse images have demonstrated the superiority of our framework with lower bitrates, higher reconstruction quality, and increased versatility towards visual analysis and content manipulation tasks.

Geometric partitioning has attracted increasing attention by its remarkable motion field description capability in the hybrid video coding framework. However, the existing geometric partitioning (GEO) scheme in Versatile Video Coding (VVC) causes a non-negligible burden for signaling the side information. Consequently, the coding efficiency is limited. In view of this, we propose a spatio-temporal correlation guided geometric partitioning (STGEO) scheme to efficiently describe the object information in the motion field of video coding. The proposed method can economize the bits consumed for side information signaling, including the partitioning mode and motion information. We firstly analyze the characteristics of partitioning mode decision and motion vector selection in a statistically-sound way. Based on the observed spatio-temporal correlation, we design a mode prediction and coding method to reduce the overhead for representing the above mentioned side information. The main idea is to predict the STGEO modes and motion candidates that have higher selection possibilities, which can guide the entropy coding, i.e., representing the predicted high-probability modes and motion candidates with fewer bits. In particular, the high-probability STGEO modes are predicted based on the edge information and history modes of adjacent STGEO-coded blocks. The corresponding motion information is represented by the index in a merge candidate list, which is adaptively inferred based on the off-line trained merge candidate selection probability. Simulation results show that the proposed approach achieves 0.95% and 1.98% bit-rate savings on average compared to VTM-8.0 without GEO for Random Access and Low-Delay B configurations, respectively.

In-loop filters have attracted increasing attention due to the remarkable noise-reduction capability in the hybrid video coding framework. However, the existing in-loop filters in Versatile Video Coding (VVC) mainly take advantage of the image local similarity. Although some non-local based in-loop filters can make up for this shortcoming, the widely-used unsupervised parameter estimation method by non-local filters limits the performance. In view of this, we propose a deformable Wiener Filter (DWF). It combines the local and non-local characteristics and supervisedly trains the filter coefficients based on the Wiener Filter theory. In the filtering process, local adjacent samples and non-local similar samples are first derived for each sample of interest. Then the to-be-filtered samples are classified into specific groups based on the patch-level noise and sample-level characteristics. Samples in each group share the same filter coefficients. After that, the local and non-local reference samples are adaptively fused based on the classification results. Finally, the filtering operation with outlier data constraints is conducted for each to-be-filtered sample. Moreover, the performance of the proposed DWF is analyzed with different reference sample derivation schemes in detail. Simulation results show that the proposed approach achieves 1.16%, 1.92%, and 2.67% bit-rate savings on average compared to the VTM-11.0 for All Intra, Random Access, and Low-Delay B configurations, respectively.

Recently, deep learning for video coding, such as deep predictive coding, deep transform coding, and deep in-loop filtering, has been an emerging research area. The coding gain of hybrid coding framework could be extensively promoted by the data-driven models. However, previous deep coding tools especially deep in-loop filtering mainly consider the performance improvement while pay less attention to the reliability, usability, and adaptivity of the networks. In this article, a nested-residual guided convolutional neural network (NR-CNN) structure with cascaded global shortcut and configurable residual blocks is proposed for in-loop filtering. By taking advantage of the correlation between different color components, we further extend the NR-CNN by utilizing luminance as textural and structural guidance for chrominance filtering, which significantly improves the filtering performance. To fully exploit the proposed network into codec integration, we subsequently introduce an efficient and adaptive framework consisting of an adaptive granularity optimization and a parallel inference pipeline for deep learning based filtering. The former contributes to the coding performance improvement through an adaptive decision-making based on rate-distortion analysis at various granularities. The latter reduces the running time of network inference. The extensive experimental results show the superiority of the proposed method, achieving 8.2%, 14.9%, and 13.2% BD-rate savings on average under random access (RA) configuration. Meanwhile, the proposed method also obtains better subjective quality.

In-loop filter is a crucial module in video coding, which can improve both subjective and object quality of reconstructed videos. In this paper, a new sample-based classification method is first proposed using features extracted from different stages of the existing in-loop filter process. Based on this method, an adaptive three-layer Quad-tree Based Sample Refinement Filter (QSRF) algorithm is designed to further improve the coding efficiency. Experimental results show that the proposed QSRF algorithm achieves 0.39%, 0.77% and 0.70% BD-rate savings for random access, lowdelay B and lowdelay P configurations compared to AVS3 reference software, respectively. Moreover, the proposed method can also improve visual quality of reconstructed videos significantly.

In the Versatile Video Coding (VVC) standard, adaptive loop filter (ALF), including Geometry transformation-based Adaptive Loop Filter (GALF) and Cross Component Adaptive Loop Filter (CCALF), plays an essential role in reducing compression artifacts. However, it also has high coding complexity and requires many picture buffer accesses in the encoder that will increase external memory access and is unfriendly to the software and hardware design. Therefore, we propose an optimized ALF framework, including the parallel design of GALF and CCALF, the adaptive parameter decision of GALF, and one-pass CCALF scheme by effectively estimating the CCALF filtering distortion without conducting filter operation. Compared to VTM-8.0, the proposed method can reduce the picture buffer access from 152 to 1 and achieve roughly 25% time-savings of the ALF module with negligible coding performance change under RA configuration. Some of the proposed methods have been adopted in the VVC reference software.

Conceptual coding has been an emerging research topic recently, which encodes natural images into disentangled conceptual representations for compression. However, the compression performance of the existing methods is still suboptimal due to the lack of comprehensive consideration of rate constraint and reconstruction quality. To this end, we propose a novel end-to-end semantic prior modeling based conceptual coding scheme towards extremely low bitrate image compression, which leverages semantic-wise deep representations as a unified prior for entropy estimation and texture synthesis. Specifically, we employ semantic segmentation maps as structural guidance for extracting deep semantic prior, which provides fine-grained texture distribution modeling for better detail construction and higher flexibility in subsequent high-level vision tasks. Moreover, a cross-channel entropy model is proposed to further exploit the inter-channel correlation of the spatially independent semantic prior, leading to more accurate entropy estimation for rate-constrained training. The proposed scheme achieves an ultra-high 1000× compression ratio, while still enjoying high visual reconstruction quality and versatility towards visual processing and analysis tasks.

Optimized for pixel fidelity metrics, images compressed by existing image codec are facing systematic challenges when used for visual analysis tasks, especially under low-bitrate coding. This paper proposes a visual analysis-motivated rate-distortion model for Versatile Video Coding (VVC) intra compression. The proposed model has two major contributions, a novel rate allocation strategy and a new distortion measurement model. We first propose the region of interest for machine (ROIM) to evaluate the degree of importance for each coding tree unit (CTU) in visual analysis. Then, a novel CTU-level bit allocation model is proposed based on ROIM and the local texture characteristics of each CTU. After an in-depth analysis of multiple distortion models, a visual analysis friendly distortion criteria is subsequently proposed by extracting deep feature of each coding unit (CU). To alleviate the problem of lacking spatial context information when calculating the distortion of each CU, we finally propose a multi-scale feature distortion (MSFD) metric using different neighboring pixels by weighting the extracted deep features in each scale. Extensive experimental results show that the proposed scheme could achieve up to 28.17% bitrate saving under the same analysis performance among several typical visual analysis tasks such as image classification, object detection, and semantic segmentation.

In this paper, a flexible and adaptive block-wise image compression architecture with restoration network is proposed to improve the rate-distortion (R-D) performance of learned image compression. In contrast to existing learned compression algorithms which perform the image based transform coding, our approach splits the input image into non-overlapped blocks and compresses each block adaptively based on their local content characteristics. To reduce the artifacts caused by the block partitioning, block fusion network is subsequently proposed to overcome the drawbacks of block-wise independent coding. Further, quantization fine-tuning training process is proposed and utilized to reduce the difference of quantization operations between the training process and testing process. Experimental results show that the proposed network architecture has the ability to obtain significant coding gain compared with existing conventional compression standards including Versatile Video Coding (VVC) and achieves comparable R-D performance with state-of-the-art learned image codec using less parameters.

Traditional image/video compression aims to reduce the transmission/storage cost with signal fidelity as high as possible. However, with the increasing demand for machine analysis and semantic monitoring in recent years, semantic fidelity rather than signal fidelity is becoming another emerging concern in image/video compression. With the recent advances in cross modal translation and generation, in this paper, we propose the cross modal compression~(CMC), a semantic compression framework for visual data, to transform the high redundant visual data~(such as image, video, etc.) into a compact, human-comprehensible domain~(such as text, sketch, semantic map, attributions, etc.), while preserving the semantic. Specifically, we first formulate the CMC problem as a rate-distortion optimization problem. Secondly, we investigate the relationship with the traditional image/video compression and the recent feature compression frameworks, showing the difference between our CMC and these prior frameworks. Then we propose a novel paradigm for CMC to demonstrate its effectiveness. The qualitative and quantitative results show that our proposed CMC can achieve encouraging reconstructed results with an ultrahigh compression ratio, showing better compression performance than the widely used JPEG baseline.

The non-local adaptive in-loop filter (NALF) for video coding has achieved significant coding gain by exploiting image non-local self-similarity (NSS) to efficiently reduce the compression artifacts. However, the intensive computation of NALF hinders its practical deployment in video standardizations. In this paper, we propose a fast NALF optimization algorithm in parallel-computing framework by leveraging the massive parallel execution resources of GPU. First, the computational complexity of original NALF is analyzed in depth, then the pipelines of computational-intensive modules are re-designed to adapt to the general-purpose GPU with more parallel-friendly consideration. Specifically, we speed up the NALF by optimizing thread allocation to maximize the parallelism degree and elaborately designing the GPU block dimension to avoid access conflict. The group-level and pixel-level parallelization for collaboratively filtering and patch matching modules are designed respectively. To reduce the cost in data transmission, the whole filtering process is implemented on GPU by taking the advantage of low data dependency in NALF. Extensive experimental results show that the proposed fast NALF optimization using GPU architecture achieves high-speeed processing while maintaining the significant coding performance of original NALF, which shows the potential of NALF in the future video coding standard.

Attacking deep learning based biometric systems has drawn more and more attention with the wide deployment of fingerprint/face/speaker recognition systems, given the fact that the neural networks are vulnerable to the adversarial examples, which have been intentionally perturbed to remain almost imperceptible for human. In this paper, we demonstrated the existence of the universal adversarial perturbations (UAPs) for the speaker recognition systems. We proposed a generative network to learn the mapping from the low-dimensional normal distribution to the UAPs subspace, then synthesize the UAPs to perturbe any input signals to spoof the well-trained speaker recognition model with high probability. Experimental results on TIMIT and LibriSpeech datasets demonstrate the effectiveness of our model.

To improve the coding performance, geometric partition (GEO) was proposed for the upcoming VVC standard. GEO provides 140 partition candidates. The index of optimal GEO mode needs to be signaled explicitly. Considering different structural characteristics of different CUs and the correlation between spatial adjacent blocks and temporal collocated blocks, we propose a GEO mode prediction strategy by constructing a Most Probable Mode (MPM) list to reduce the overhead of GEO index and improve coding efficiency. Based on the observation of the high correlation between the partition mode and object boundaries, an edge-directed geometric partition scheme is proposed to construct the MPM list according to spatio-temporal edge information. The proposed method provides an objective BD-rate gain of 0.58% and 1.00% on average for RA and LDB configurations compared to VTM-6.0. Besides, it also promotes the visual quality of object boundaries.

In recent years, the image and video coding technologies have advanced by leaps and bounds. However, due to the popularization of image and video acquisition devices, the growth rate of image and video data is far beyond the improvement of the compression ratio. In particular, it has been widely recognized that there are increasing challenges of pursuing further coding performance improvement within the traditional hybrid coding framework. Deep convolution neural network which makes the neural network resurge in recent years and has achieved great success in both artificial intelligent and signal processing fields, also provides a novel and promising solution for image and video compression. In this paper, we provide a systematic, comprehensive and up-to-date review of neural network-based image and video compression techniques. The evolution and development of neural network-based compression methodologies are introduced for images and video respectively. More specifically, the cutting-edge video coding techniques by leveraging deep learning and HEVC framework are presented and discussed, which promote the state-of-the-art video coding performance substantially. Moreover, the end-to-end image and video coding frameworks based on neural networks are also reviewed, revealing interesting explorations on next generation image and video coding frameworks/standards. The most significant research works on the image and video coding related topics using neural networks are highlighted, and future trends are also envisioned. In particular, the joint compression on semantic and visual information is tentatively explored to formulate high efficiency signal representation structure for both human vision and machine vision, which are the two dominant signal receptors in the age of artificial intelligence.

Deep learning based video coding tools development has been an emerging topic recently. In this paper, we propose a novel deep convolutional neural network (CNN) based in-loop filter algorithm for the third generation of Audio Video Coding Standard (AVS3). Specifically, we first introduce a residual block based CNN model with global identity connection for the luminance in-loop filter to replace conventional rule-based algorithms in AVS3. Subsequently, the reconstructed luminance channel is deployed as textural and structural guidance for chrominance filtering. The corresponding syntax elements are also designed for the CNN based in-loop filtering. In addition, we build a large scale database for the learning based in-loop filtering algorithm. Experimental results show that our method achieves on average 7.5%, 16.9% and 18.6% BD-rate reduction under all intra (AI) configuration on common test sequences. In particular, the performance for 4K videos is 6.4%, 15.5% and 17.5% respectively. Moreover, under random access (RA) configuration, the proposed method brings 3.3%, 14.4%, and 13.6% BD-rate reduction separately.

The paper presents a systematic and comprehensive overview of the third generation of Audio Video Standard (AVS) in China. The AVS standards continuously attract extensive attention both domestically and worldwide along with the industrialization of AVS2 standard and the broadcasting of China Central Television (CCTV) 4K channel. Currently, AVS3 has adopted many novel coding techniques including block partitioning structure, intra/inter and transform coding tools. The compression performance has been significantly improved that the latest version of AVS3 obtains about 24.3% and 26.88% bit-rate reduction against AVS2 and HEVC respectively under 4K resolution sequences.

In this paper, we provide a systematical overview and analysis on the joint feature and texture representation framework, which aims to smartly and coherently represent the visual information with the front-end intelligence in the scenario of video big data applications. In particular, we first demonstrate the advantages of the joint compression framework in terms of both reconstruction quality and analysis accuracy. Subsequently, the interactions between visual feature and texture in the compression process are further illustrated. Finally, the future joint coding scheme by incorporating the deep learning features is envisioned, and future challenges toward seamless and unified joint compression are discussed. The joint compression framework, which bridges the gap between visual analysis and signal-level representation, is expected to contribute to a series of applications, such as video surveillance and autonomous driving.

Recently, convolutional neural network (CNN) has attracted tremendous attention and has achieved great success in many image processing tasks. In this paper, we focus on CNN technology combined with image restoration to facilitate video coding performance and propose the content-aware CNN based in-loop filtering for high-efficiency video coding (HEVC). In particular, we quantitatively analyze the structure of the proposed CNN model from multiple dimensions to make the model interpretable and optimal for CNN-based loop filtering. More specifically, each coding tree unit (CTU) is treated as an independent region for processing, such that the proposed content-aware multimodel filtering mechanism is realized by the restoration of different regions with different CNN models under the guidance of the discriminative network. To adapt the image content, the discriminative neural network is learned to analyze the content characteristics of each region for the adaptive selection of the deep learning model. The CTU level control is also enabled in the sense of rate-distortion optimization. To learn the CNN model, an iterative training method is proposed by simultaneously labeling filter categories at the CTU level and fine-tuning the CNN model parameters. The CNN based in-loop filter is implemented after sample adaptive offset in HEVC, and extensive experiments show that the proposed approach significantly improves the coding performance and achieves up to 10.0% bit-rate reduction. On average, 4.1%, 6.0%, 4.7%, and 6.0% bit-rate reduction can be obtained under all intra, low delay, low delay P, and random access configurations, respectively.

HEVC is the latest video coding standard, in which inter prediction plays an important role to reduce the temporal redundancy. The accuracy of inter prediction is limited since only temporal information is used in conventional algorithms. In this paper, we propose a neural network based inter prediction algorithm for HEVC by using the spatial-temporal information. In the proposed algorithm, we first design a neural network architecture consisting of a fully connected network (FCN) and a convolutional neural network (CNN). Then the spatial neighboring pixels and the temporal neighboring pixels are inputted into FCN. The output of FCN and the prediction of current block are inputted into CNN, which will results the more accurate prediction of current block. Experimental results demonstrate that the proposed method can achieve average 1.7% (up to 8.6%) BD-rate reduction in low delay P test condition compared to HM 16.9.

In this paper, we propose a novel light field image compression scheme by exploiting the intrinsic similarity of light field images with deep learning. In particular, instead of conveying all LF sub-views, only sparsely sampled LF sub-views are compressed and the remaining sub-views are reconstructed from the coded sub-views in the neighbourhood with convolutional neural network (CNN). To jointly suppress the artifacts induced in compression and reconstruct the un-coded views with high geometric accuracy, a multi-view joint enhancement network is introduced to improve the coding performance. Extensive experiments show the superior compression performance of our scheme compared with the state-of-the-art methods.

In order to compensate the shortcomings of existing in-loop filters only based on local correlation in video coding standards, many non-local based loop filters with high coding performance and computational complexity are proposed. In this paper, we propose a fast block matching algorithm, adaptive two-step block matching algorithm, based on our previous work, structure-driven adaptive non-local filter (SANF) which is computationally intensive because of the high complexity of block matching and singular value decomposition (SVD). Our proposed algorithm based on image spatial statistical characteristics utilizes fixed template to select adaptive number of similar blocks according to image content, which can reduce up to 75.2% search candidates compared to exhaustive search in SANF and the adaptive determination strategy can remove blocks with less relation to reference block in similar block group which have little help for compression performance, and the remove of them can reduce the computational complexity of SVD. Our proposed optimization algorithm can save encoding and decoding time significantly with negligible performance loss, which achieves 70.7%, 84.4%, 80.82% and 81.95% decoding time saving with only 0.13%, 0.05%, 0.13% and 0.15% increases of BD-rate for AI, RA, LDB and LDP configurations, respectively compared to original SANF in JEM-7.0.

In this paper, we explore the structure of light field (LF) and efficiently improve the performance of the pseudo sequence based lenslet image compression, by optimized sub-view rearrangement, enhanced illumination compensation and adaptive reconstruction filtering. First, the decomposed sub-view images are rearranged into a pseudo sequence according to our optimized scan order based on sub-view correlation. Second, the generated pseudo sequence is compressed with JEM codec in which we enhance the illumination compensation by adaptively selecting reference pixels used in parameter derivation. Finally, to reduce the distortions in lenslet image decomposition and reconstruction, we propose an lenslet reconstruction method by applying adaptive filters to the reconstructed lenslet images to compensate the reconstruction errors. Each filter is derived by minimizing the distortions between the original and reconstructed pixels with same geolocation in the lenslet image. Extensive experimental results show that the proposed method achieves up to 53.7% bit rate reduction over HEVC intra coding and 34.8% over JEM intra coding in terms of BDBR.

Deep learning has demonstrated tremendous break through in the area of image/video processing. In this paper, a spatial-temporal residue network (STResNet) based in-loop filter is proposed to suppress visual artifacts such as blocking, ringing in video coding. Specifically, the spatial and temporal information is jointly exploited by taking both current block and co-located block in reference frame into consideration during the processing of in-loop filter. The architecture of STResNet only consists of four convolution layers which shows hospitality to memory and coding complexity. Moreover, to fully adapt the input content and improve the performance of the proposed in-loop filter, coding tree unit (CTU) level control flag is applied in the sense of rate-distortion optimization. Extensive experimental results show that our scheme provides up to 5.1% bit-rate reduction compared to the state-of-the-art video coding standard.

Deblocking filter (DF) Is High Efficiency Video Coding (HEVC) is Only Applied to all Samples Adjacent to prediction units (PU), or transform units (TU), which actually exists two issues. The first one is that DF in HEVC does not fully exploit nonlocal similarity structure information in video. The second one is that DF is HEVC does not consider the inside pixels, which often suffer from quantization distrotion. To alleviate these issues, in this paper, a structure-driven adaptive non-local filter (SANF) Is Proposed By Simultaneously Enforcing The Intrinsic Local Sparsity And The Non-Local Self-Similarity Of Each Frame. Not only SANF deals with the boundary pixels, but also the inside area, which is able to effectively reduce block artifacts while enhancing the quality of the deblocked frames. Applying SANF to luma and chroma components after DF, simulation results demonstrate that the proposed SANF can save BD-rate reduction up to 10.3% with ALF off. For luma component, SANF achieves 4.1%. 3.3%, 4.4% BD-rate saving for all intra, low delay B and random access configurations, respectively with ALF off. furthermore, the performance with ALF on is also discussed.

In this contribution, a novel image quality enhancement algorithm based on convolutional network is proposed for low bit rate image compression. Specifically, a downsample procedure is performed to generate lower resolution image for low bit rate compression. While the decoder side, upsample is to be performed firstly to the original resolution. Image quality is further enhanced by the proposed convolutional deep network. In particular, an optional image quality improvement network can be utilized for further enhancement after the first network. With the help of deep network, more detailed and high-frequency information can be recovered while maintaining the consistency of contour area, leading to better visual quality. Another benefit of this approach lies in that the proposed approach is fully compatible with all third-party image codec pipeline. Experimental result shows that the proposed scheme significantly outperforms JPEG in low bit rate image compression.

The deblocking filtering (DF) in HEVC is only applied to the boundaries between the coding units, prediction units, or transform units, which actually exists two issues. On one hand, the simple DF in HEVC does not fully exploit structure information in video. On the other hand, DF in HEVC does not consider the inside areas, which often suffers from quantization distortion. To alleviate the above issues, in this paper, a non-local structure-based filter (NLSF) is proposed by simultaneously enforcing the intrinsic local sparsity and the non-local self-similarity of each frame in video. NLSF not only deals with the boundaries, but also deals with the inside areas, which is able to effectively reduce the artifacts while enhance the quality of the deblocking frames. Experimental results demonstrate that, compared with the original HEVC reference encoder implementation in AI configuration, the proposed NLSF can achieve up to 7.3% BD-rate saving by substituting for DF in HEVC.