Training deep neural networks (DNNs) typically requires large-scale datasets, which poses substantial challenges related to computing resources and storage. Dataset Quantization (DQ) was introduced to compress large datasets into smaller subsets for training various neural networks. However, DQ lacks consideration of sample importance and diversity, which may result in excluding crucial samples or insufficiently representing the dataset, thereby limiting model generalization. To resolve this, we propose Dataset Quantization Augmentation (DQA), an enhanced framework that integrates image complexity into dataset compression and dynamically applies data augmentation techniques based on sample complexity. By selecting the least complex samples and applying data augmentation techniques (e.g., CutBlur, CutMix, Cutout, geometric transformations), DQA enhances dataset diversity and boosts model performance. Experimental results demonstrate that DQA outperforms the original DQ method and other Dataset compression methods. On CIFAR-10, DQA achieves a 3.64% accuracy improvement over DQ with only 10% of the dataset. For HDR reconstruction using the HAT model on Flickr2K, DQA achieves a PSNR 0.1761 higher and an SSIM 0.132 higher than DQ at 30% compression. These results highlight DQA’s versatility and effectiveness in both high-level and low-level vision tasks, making it a promising approach for dataset optimization.

To reconstruct high dynamic range (HDR) video from alternating exposed low dynamic range (LDR) frames, the key is to address the misalignment and imprecise fusion caused by information loss and noise in ill-exposed regions. Following a coarse-to-fine manner, a Multi-level Spatial-Temporal feature aggregation and alignment-based Selective Residual Dense Propagation Network (MSTSRDPNet) is proposed. The Multi-level Spatial-Temporal aggregation extracts spatial-temporal features and aggregates them to mitigate information loss for fusion. The alignment-based Selective Residual Dense Propagation module reconstructs the aligned feature by using channel attention to redistribute feature weights while leveraging residual dense connections for information propagation. Experiments show that the proposed MSTSRDPNet outperforms all conventional methods on the synthetic dataset with PSNR-T, HDR-VQM, and HDR-VDP-2 scores of 44.64 dB, 86.83, and 73.9.

To generate ghost-free high dynamic range (HDR) images by merging multiple differently exposed low dynamic range (LDR) images, the key is to handle ill-exposed areas in the input LDR images and misalignment among them. In this paper, a Pyramid Spatial Feature Transform and shared-offsets Deformable convolutional Network (PSFTDNet) is proposed to achieve this target. The pyramid spatial feature transform module tackles ill-exposed areas, which modulates the features to exploit complementary information of them in a coarse-to-fine manner. The shared-offsets deformable alignment handles misalignment among the input images, which applies the offsets used to align exposure-aligned images to align all the features. Experiments on the NTIRE HDR challenge dataset and Kalantari dataset show that the proposed PSFTDNet outperforms all the conventional methods with PSNR-L scores of 42.31 dB and 41.54 dB, and PSNR-T scores of 34.78 dB and 43.56 dB.

Mainstream image manipulation-based data augmentation methods (e.g., CutMix) undermine the integrity of extracted features, which leads to limited effect for pixel-level image processing tasks. In this paper, a local area-based mixed data augmentation method called HDR-LMDA for HDR video reconstruction is proposed. Within it, the local exposure augmentation (LEA) applies different exposures to original LDR inputs among different regions, while the local RGB permutation (LRP) shuffles the color channels of the random patch instead of the entire frame. By mixing both two operations, the model is forced to learn how to apply concise ill-exposure recovery and color processing within the same training process. Experiments demonstrate that HDR-LMDA achieves a better PSNR-T boost of 0.93dB, compared with conventional works under the same conditions.

Abstract High‐dynamic‐range image (HDRI) displays scenes as vivid as the real scenes. HDRI can be reconstructed by fusing a set of bracketed‐exposure low‐dynamic‐range images (LDRI). For the reconstruction, many works succeed in removing the ghost artefacts caused by moving objects. The critical issue is reconstructing the areas which are saturated due to bad exposure and occluded due to motion with no ghost artefacts. To overcome this issue, this paper proposes texture and exposure awareness based refill. The proposed work first locates the saturated and occluded areas existing in input image set, then refills background textures or patches containing rough exposure and colour information into located areas. Proposed work can be integrated with multiple existing ghost removal works to improve the reconstruction result. Experimental results show that proposed work removes the ghost artefacts caused by saturated and occluded areas in subjective evaluation. For the objective evaluation, the proposed work improves the HDR‐VDP‐2 evaluation result for multiple conventional works by 1.33% on average.

Ghost-like artifact caused by ill-exposed and motion areas is one of the most challenging problems in high dynamic range (HDR) image reconstruction. When the motion range is small, previous methods based on optical flow or patch-match can suppress ghost-like artifacts by first aligning input images before merging them. However, they are not robust enough and still produce artifacts for challenging scenes where large foreground motions exist. To this end, we propose a deep network with an attention module and motion-emphasized loss function to produce ghost-free HDR images. In the attention module, we use the channel and spatial attention to guide the network to emphasize important components such as motion and saturated areas automatically. To be robust to images with different resolutions and objects with distinct scales, we adopt the selective kernel network as the basic framework for channel attention. In addition to the attention module, the motion-emphasized loss function based on the motion and ill-exposed areas mask is designed to help the network reconstruct motion areas. Experiments on the public dataset indicate that the proposed SK-AHDRNet produces ghost-free results where detail in ill-exposed areas is well recovered. The proposed method scores 43.17 with PSNR metric and 61.02 with HDR-VDP-2 metric on test which outperforms all conventional works. According to quantitative and qualitative evaluations, the proposed method can achieve state-of-the-art performance.

A growing number of new technologies are used in test development. Among them, automatic test generation, a promising technology to improve the efficiency of unit testing, currently performs not satisfactory in practice. Test recommendation, like code recommendation, is another feasible technology for supporting efficient unit testing and gets increasing attention. In this paper, we develop a novel system, namely HomoTR, which implements online test recommendations by measuring the homology of two methods. If the new method under test shares homology with an existing method that has test cases, HomoTR will recommend the test cases to the new method. The preliminary experiments show that HomoTR can quickly and effectively recommend test cases to help the developers improve the testing efficiency. Besides, HomoTR has been integrated into the MoocTest platform successfully, so it can also execute the recommended test cases automatically and visualize the testing results (e.g., Branch Coverage) friendly to help developers understand the process of testing. The demo video of HomoTR can be found at https://youtu.be/_227EfcUbus.

High-dynamic-range imaging (HDRI) techniques are proposed to extend the dynamic range of captured images against sensor limitation. The key issue of multi-exposure fusion in HDRI is removing ghost artifacts caused by motion of moving objects and handheld cameras. This paper proposes a ghost-free HDRI algorithm based on visual salience and stack extension. To improve the accuracy of ghost areas detection, visual salience based bilateral motion detection is introduced to measure image differences. For exposure fusion, the proposed algorithm reduces brightness discontinuity and enhances details by stack extension, and rejects the information of ghost areas to avoid artifacts via fusion masks. Experiment results show that the proposed algorithm can remove ghost artifacts accurately for both static and handheld cameras, remain robust to scenes with complex motion and keep low complexity over recent advances including patch based method and rank minimization based method by 20.4% and 63.6% time savings on average.

In H.264/AVC based integer motion estimation engine, fixed architectures based on full pixel or direct sub-sampling pattern are widely used for HDTV application. However, these architectures suffer from either high complexity or quality loss problems. In this paper, an adaptive sub-sampling based reconfigurable architecture is given out. Firstly, by executing pixel difference analysis, the adaptive sub-sampling scheme which uses three hardware friendly patterns is applied on homogeneous macroblock (MB). Secondly, the related architecture introduces one more pipeline stage to build up configurable partial SAD values so that system performance is enhanced. Thirdly, a two-level pixel data organization scheme is proposed to solve data reuse and hardware utilization problems caused by adaptive algorithm. Moreover, one cross based SAD generation structure is introduced to achieve adaptive output results with less hardware cost. Experimental results show that, the proposed architecture can averagely save 61.71% clock cycles and accomplish twice or four times processing capability for homogeneous MBs. The maximum clock frequency is 208MHz under the TSMC 0.18um technology in worst case conditions(1.62V, 125 C).

One fast inter mode decision algorithm is proposed in this paper. The whole algorithm is convoluted with block matching process. Firstly, before ME process, by exploiting spatial and temporal information, a skip mode early detection algorithm is proposed. Also, in this stage, edge gradient is used to filter out unpromising modes. Secondly, during the ME stage, the original search window is separated into several layers and our fast decision scheme works with motion information of each layer. Moreover, before stepping into small modes, the distribution of SAD (sum-of-absolute-difference) and RD (rate distortion) costs of big modes are analyzed in an early stage to accelerate the inter mode decision process. Experiments show that our algorithm can achieve a speed-up factor of up to 66.0% with trivial bit increment and quality degradation.

In this paper, we contribute a configurable SAD tree architecture based on adaptive subsampling scheme. Firstly, by further exploiting the spatial feature, the integer motion estimation process is greatly sped up. Secondly, the conventional partial sum of absolute difference (SAD) based pipeline structure is optimized into configurable SAD oriented way, which enhances the performance and solve the data reuse problem caused by adaptive scheme in the architecture level. Moreover, a cross reuse and compressor tree based circuit level optimization is introduced and 6.56% hardware cost is reduced. Experiments show that our design can averagely achieve 42.23% saving in processing cycles compared with previous design. With 323 k gates at about 144.8 MHz, our design can achieve real-time encoding of HDTV 1088 p@30 fps.

One Super Hi-Vision (SHV) 4k times 4k @60 fps fractional motion estimation (FME) engine is proposed in our paper. Firstly, two complexity reduction schemes are proposed in the algorithm level. By analyzing the integer motion cost, 48% clock cycle is saved based on our mode pre-filtering scheme. By further check the motion cost of neighboring search points, our directional one-pass scheme can achieve reduction of 50% clock cycle and 36% hardware cost. Secondly, in the hardware level, two parallel improved schemes namely 16-Pel interpolation and MB-parallel processing are given out. Thirdly, one unified pixel block loading scheme is proposed. About 28.67% to 80.68% pixels are reused and the related memory access is saved. Furthermore, one parity pixel organization scheme is proposed to solve memory access conflict of MB-parallel processing. By using TSMC 0.18 mum technology in worst work conditions (1.62 V, 125degC), our FME engine can achieve real-time processing for SHV 4ktimes4k@60fps with 976.5 k gates hardware.

This paper presents an on bit allocation and Lagrange multiplier (lambdaMODE) adjustment for H.264 rate control. To enhance the precision of the complexity estimation and bit allocation, a frequency-domain parameter named mean-absolute-transform-difference (MATD) is adopted to represent frame and macroblock (MB) residual complexity. Second, the MATD ratio is utilized to enhance the accuracy of frame layer bit prediction. Then, by considering the bit usage status of whole sequence, a measurements combining forward and backward bit analysis is proposed to optimize the lambdaMODEon frame layer. On MB layer, bits are allocated by proposed remaining complexity analysis. The computed quantization parameter (QP) is further adjusted according to predicted MB texture bits. Simulation results verify the performance of the proposed algorithm. Compared with the recommended rate control in H.264 reference software JM13.2, the PSNR improvement is up to 1.13 dB by using our algorithm.

The search range (SR) parameter plays an important role in motion estimation (ME) for video coding. Adaptively adjusting SR according to the information given by previously encoded syntax element, also known as adaptive search range (ASR) algorithm, can efficiently reduce the computational complexity of ME. Compared with heuristic search pattern (HSP) algorithms like diamond/hexagon search, ASR algorithms are more fundamental, flexible and hardware-oriented. This paper although starts with a comparison between HSP and ASP algorithms which is followed by a proposed ASR algorithm with experimental results, however more likely intends to contribute several novel perspectives to this research area.

8 times 8 transform has been introduced in H.264psilas high profile to improve the video quality. After transform and quantization, if all the coefficients of the blockpsilas residue data become zero, this block is called all-zero block (AZB). Many 4 times 4 transform AZB early detection algorithms have been proposed to skip transform and quantization process. In this paper, after theoretical analysis performed for the sufficient condition of 8 times 8 AZB detection, sum of absolute differences(SAD) and sum of absolute transformed(SATD) based 8 times 8 AZB detection algorithms are proposed. Experimental results show that the proposed algorithms achieve major improvement of computation reduction from 0.95% to 79.48% for 720p sequences. The computation reduction increases as QP increases.