Built on deep networks, end-to-end optimized image compression has made impressive progress in the past few years. Previous studies usually adopt a compressive auto-encoder, where the encoder part first converts image into latent features, and then quantizes the features before encoding them into bits. Both the conversion and the quantization incur information loss, resulting in a difficulty to optimally achieve arbitrary compression ratio. We propose iWave++ as a new end-to-end optimized image compression scheme, in which iWave, a trained wavelet-like transform, converts images into coefficients without any information loss. Then the coefficients are optionally quantized and encoded into bits. Different from the previous schemes, iWave++ is versatile: a single model supports both lossless and lossy compression, and also achieves arbitrary compression ratio by simply adjusting the quantization scale. iWave++ also features a carefully designed entropy coding engine to encode the coefficients progressively, and a de-quantization module for lossy compression. Experimental results show that lossy iWave++ achieves state-of-the-art compression efficiency compared with deep network-based methods; on the Kodak dataset, lossy iWave++ leads to 17.34 percent bits saving over BPG; lossless iWave++ achieves comparable or better performance than FLIF. Our code and models are available at https://github.com/mahaichuan/Versatile-Image-Compression.

We address the requirement of image coding for joint human-machine vision, i.e., the decoded image serves both human observation and machine analysis/understanding. Previously, human vision and machine vision have been extensively studied by image (signal) compression and (image) feature compression, respectively. Recently, for joint human-machine vision, several studies have been devoted to joint compression of images and features, but the correlation between images and features is still unclear. We identify the deep network as a powerful toolkit for generating structural image representations. From the perspective of information theory, the deep features of an image naturally form an entropy decreasing series: a scalable bitstream is achieved by compressing the features backward from a deeper layer to a shallower layer until culminating with the image signal. Moreover, we can obtain learned representations by training the deep network for a given semantic analysis task or multiple tasks and acquire deep features that are related to semantics. With the learned structural representations, we propose SSSIC, a framework to obtain an embedded bitstream that can be either partially decoded for semantic analysis or fully decoded for human vision. We implement an exemplar SSSIC scheme using coarse-to-fine image classification as the driven semantic analysis task. We also extend the scheme for object detection and instance segmentation tasks. The experimental results demonstrate the effectiveness of the proposed SSSIC framework and establish that the exemplar scheme achieves higher compression efficiency than separate compression of images and features.

In this paper, we consider a novel image coding paradigm, termed semantically scalable coding. In the new paradigm, coded bitstream serves for multiple different semantic analysis tasks, and different tasks require different semantic granularities of the image. Thus, the bitstream is designed to be scalable in the sense that progressive decoding of the bitstream provides coarse-to-fine semantic granularities. As a concrete example, we consider the task of coarse-grained and fine-grained image classification. We present a method to compress the multiple deep feature maps that are intermediate representations of an image passing a trained deep network. The deep-layer feature maps can serve for coarse-grained image classification while the shallow-layer feature maps can serve for fine-grained image classification. Experimental results demonstrate the feasibility of the proposed method, as well as the advantage of the semantically scalable coding paradigm.

We propose an image coding scheme that compresses image into semantically scalable bitstream using deep neural networks. This scheme is expected to support intelligent analysis when the bitstream is partially decoded, as well as high-fidelity reconstruction of image when the bitstream is completely decoded. We implement such a semantically scalable image coding scheme based on semantic map. In the proposed scheme, the original image is firstly semantically segmented and the semantic map is compressed as the base layer. Then, the original image is segmented into several individual objects according to the semantic map, and each object is coded separately. A recurrent neural network-based encoder is used to compress these objects at several quality levels. At the decoder side, the semantic map can be directly applied for intelligent analysis. A generative adversarial network is used to synthesize a rough image using the semantic map. If user is interested in a certain object, more bits can be transmitted to enhance the quality of the object. Experimental results show that the proposed method achieves comparable compression performance with JPEG2000 at high bit rates, while facilitates intelligent analysis at low bit rates.

In video-based point cloud compression (V-PCC), occupancy map video is utilized to indicate whether a 2-D pixel corresponds to a valid 3-D point or not. In the current design of V-PCC, the occupancy map video is directly compressed losslessly with High Efficiency Video Coding (HEVC). However, the coding tools in HEVC are specifically designed for natural images, thus unsuitable for the occupancy map. In this paper, we present a novel quadtree-based scheme for lossless occupancy map coding. In this scheme, the occupancy map is firstly divided into several coding tree units (CTUs). Then, the CTU is divided into coding units (CUs) recursively using a quadtree. The quadtree partition is terminated when one of the three conditions is satisfied. Firstly, all the pixels have the same value. Secondly, the pixels in the CU only have two kinds of values and they can be separated by a continuous edge whose endpoints lie on the side of the CU. The continuous edge is then coded using chain code. Thirdly, the CU reaches the minimum size. This scheme simplifies the design of block partitioning in HEVC and designs simpler yet more effective coding tools. Experimental results show significant reduction of bit-rate and complexity compared with the occupancy map coding scheme in V-PCC. In addition, this scheme is also very efficient to compress the semantic map.

In video action recognition, motion is a very crucial clue, which is usually represented by optical flow. However, optical flow is computationally expensive to obtain, which becomes the bottleneck for the efficiency of traditional action recognition algorithms. In this article, we propose a network called MV2Flow to learn motion representation efficiently from the signals in the compressed domain. To learn the network, three losses are defined. First, we select the classical TV-L1 flow as proxy ground truth to guide the learning. Besides, an unsupervised image reconstruction loss is proposed to further refine it. Moreover, toward the task of action recognition, the above two losses are combined with a motion content loss. To evaluate our approach, extensive experiments on two benchmark datasets UCF-101 and HMDB-51 are conducted. The motion representation generated with our MV2Flow has shown comparable classification performance on action recognition with TV-L1 flow, while operating at an over 200× faster speed. Based on our MV2Flow and 2D-CNN-based network, we have achieved state-of-the-art performance in the compressed domain. With 3D-CNN-based network, we also achieve comparable accuracy with higher inference speed than methods in the decoded domain setting.

Fractional-pixel motion compensation (MC) improves the efficiency of inter prediction and has been utilized extensively in video coding standards. The traditional methods of fractional-pixel MC usually follow the approach of interpolation, i.e., they adopt different kinds of filters, either fixed or adaptive, to interpolate fractional-pixel values from integer-pixel values in a reference picture. Different from the interpolation approach, in this paper, we formulate the fractional-pixel MC as an inter-picture regression problem, which is to predict the pixel values of the current to-be-coded picture from the integer-pixel values of a reference picture, given a fractional-pixel motion vector that relates the two pictures. We then propose to adopt convolutional neural network (CNN) models to approach the regression problem, inspired by the recent advances of CNN. Accordingly, we propose fractional-pixel reference generation CNN (FRCNN) for both uni-directional and bi-directional MC in video coding. We further investigate how to train FRCNN by using encoded video sequences, and empirically study the effect of different training data and different CNN structures. Moreover, we propose to integrate FRCNN into the high efficiency video coding (HEVC) scheme, and perform a comprehensive set of experiments to evaluate the effectiveness of FRCNN. The experimental results show that our proposed FRCNN achieves on average 3.9%, 2.7%, and 1.3% bits saving compared with HEVC, under low-delay P, low-delay B, and random-access configurations, respectively.

Motion compensation with fractional motion vector has been widely utilized in the video coding standards. The fractional samples are usually generated by fractional interpolation filters. Traditional interpolation filters are usually designed based on the signal processing theory with the assumption of band-limited signal, which cannot effectively capture the non-stationary property of video content and cannot adapt to the variety of video quality. In this paper, we reveal an intuitive property of the fractional interpolation problem, named invertibility. That is, the fractional interpolation filters should not only generate fractional samples from integer samples but also recover the integer samples from the fractional samples in an invertible manner. We prove in theory that the invertibility in the spatial domain is equivalent to the constant magnitude in the Fourier transform domain. Driven by the invertibility, we then develop a learning-based method to solve the fractional interpolation problem. Inspired by the advances of convolutional neural network (CNN), we propose to establish an end-to-end scheme using CNN to train invertibility-driven interpolation filter (InvIF). Different from the previous learning-based methods, the proposed training scheme does not need hand-crafted "ground truth" of fractional samples. The proposed InvIF is integrated into high efficiency video coding (HEVC), and extensive experiments are conducted to verify its effectiveness. The experimental results show that the proposed method can achieve on average 4.7% and 3.6% BD-rate reduction compared with the HEVC anchor, under low-delay-B and random-access configurations, respectively.

Fractional-pixel interpolation has been widely used in the modern video coding standards to improve the accuracy of motion compensated prediction. Traditional interpolation filters are designed based on the signal processing theory. However, video signal is non-stationary, making the traditional methods less effective. In this paper, we reveal that the interpolation filter can not only generate the fractional pixels from the integer pixels, but also reconstruct the integer pixels from the fractional ones. This property is called invertibility. Inspired by the invertibility of fractional-pixel interpolation, we propose an end-to-end scheme based on convolutional neural network (CNN) to derive the invertible interpolation filter, termed CNNInvIF. CNNlnvIF does not need the “ground-truth” of fractional pixels for training. Experimental results show that the proposed CNNInvIF can achieve up to 4.6% and on average 2.2% BD-rate reduction than HEVC under the low-delay P configuration.

Motion compensation is a fundamental technology in video coding to remove the temporal redundancy between video frames. To further improve the coding efficiency, sub-pel motion compensation has been utilized, which requires interpolation of fractional samples. The video coding standards usually adopt fixed interpolation filters that are derived from the signal processing theory. However, as video signal is not stationary, the fixed interpolation filters may turn out less efficient. Inspired by the great success of convolutional neural network (CNN) in computer vision, we propose to design a CNN-based interpolation filter (CNNIF) for video coding. Different from previous studies, one difficulty for training CNNIF is the lack of ground-truth since the fractional samples are actually not available. Our solution for this problem is to derive the “ground-truth” of fractional samples by smoothing high-resolution images, which is verified to be effective by the conducted experiments. Compared to the fixed half-pel interpolation filter for luma in High Efficiency Video Coding (HEVC), our proposed CNNIF achieves up to 3.2% and on average 0.9% BD-rate reduction under low-delay P configuration.

This paper presents diagonal motion partitions (DMP) for inter prediction in HEVC. In addition to the square and rectangular partitions, we propose to add diagonal shaped partitions to match different motion parts with oblique boundaries. Considering the overlap of pixels along the partition boundaries, the calculation of sum of absolute differences (SAD) and the motion compensation for the pixels on the boundaries are weighted. Besides, the residues of a diagonal prediction unit (PU) are augmented to form a rectangular one, so as to perform Hadamard transform on the residues. We also revise the advanced motion vector prediction (AMVP) and merge candidates based on the diagonal motion partitions. Experimental results show that on average 0.8%-1.0% BD-rate reduction can be achieved by DMP.