We present a standard-compatible video coding scheme with end-to-end optimized neural wrapper over standard video codecs that achieves significant rate-distortion (R-D) performance gains and is still efficient in decoding. We train a pair of pre- and post-processor using a differential JPEG proxy. The pre-processor applies a learned transform to the video and downsamples the video by a factor of 2. It generates a bottleneck video to be coded by a standard codec as a YUV sequence. The post-processor takes the decoded bottleneck video, does the inverse transform, and upsamples it to the original resolution. We follow the design in [1] , where we configure downsample using a layer of strided convolution. We optimize the post-processor for efficiency by replacing convolutions with kernel size larger than 1×1 to depth-wise convolutions [2] .

We propose a convolutional neural network-based coefficients prediction (CNNCP) method for intra-predicted residues in the High Efficiency Video Coding (HEVC) standard. In HEVC, discrete cosine transform (DCT) or discrete sine transform (DST) is adopted to convert the intra-predicted residues in the spatial domain into coefficients in the frequency domain. Each coefficient is scalar quantized and entropy coded into the bitstream. As DCT or DST is non-optimal linear transform, there still exist linear and non-linear correlations among different coefficients after the transform. In addition, there exist coefficients' correlations between current block and neighboring blocks, as these correlations cannot be completely exploited in the intra prediction. We thus propose to perform coefficients prediction to further reduce the redundancy among coefficients. The coefficients prediction is achieved using trained convolutional neural networks (CNNs), as CNNs can build complex relationship between input and output by training with a lot of data. In addition, a flag that signals whether to perform coefficients prediction or not at the coding unit level is transmitted to decoder. The proposed CNNCP method is implemented upon the HEVC reference software. Experimental results show that the proposed method achieves on average 1.8%, 4.1%, and 4.5% BD-rate reduction ratios in Y, U, V, respectively, compared with the HEVC baseline in all-intra configuration. In particular, the average BD-rate reduction ratios for 4K test sequences are 2.9%, 6.5%, and 6.6%.

360-degree video has become popular in recent years with the advances in virtual reality (VR) and augmented reality (AR) technologies and has been rapidly commercialized. To provide viewers with an immersive experience, 360-degree video requires higher resolution and much higher bandwidth compared with conventional 2D video. In a typical 360-degree video compression and delivery framework, the stitched input 360-degree videos, represented in a native projection format, e.g., equirectangular (ERP), are converted into another projection format, e.g., cubemap (CMP), octahedron (OHP), etc. and frame packed before being fed into existing video codecs. The intermediate projection format is important and would potentially improve the representation efficiency and coding performance. Among all the projection solutions, CMP is very popular and has been widely used in the computer graphics community. The intrinsic rectilinear properties of the CMP format are advantageous for the translational motion model in the modern codec architecture. However, in the CMP representation, the samples on the sphere are not evenly distributed within the faces, resulting in a higher density near the face boundaries and a lower density near the face center. Such non-uniform sampling scheme penalizes the video representation efficiency and degrades the coding performance. Adjusted cubemap projection (ACP) was proposed to address such non-uniform sampling by introducing transform functions to improve the sampling uniformity. However, the transform function parameters in ACP are fixed regardless of the content inside each cube face. In this paper, a generalized hybrid cubemap projection (HCP) is proposed to improve the 360-degree video coding efficiency beyond ACP. HCP is defined by a pair of forward transform and inverse transform functions with a pair of horizontal and vertical transform parameters per cube face. The encoder can choose the optimal sampling for each face by adjusting the parameters in the horizontal and vertical directions based on the 360-degree video content characteristics inside each cube face. In order to maintain the boundary continuities between two neighboring faces, in a 3x2 packing layout, vertical parameter constraints are imposed such that faces in each face-row have the same vertical parameters. The HCP parameters are chosen to minimize the end-to-end weighted conversion error and determined using iterative search between the horizontal and the vertical directions. Significant changes in HCP parameter values can cause drastic change in sampling distribution, and may affect the inter-picture coding efficiency. Therefore, an efficient HCP parameter estimation algorithm is proposed to achieve a better trade-off between the temporal sampling adaptation and the inter-picture prediction efficiency by reducing the temporal variation of HCP parameters. The proposed HCP parameter search algorithm reduces the computational complexity by 5x compared to the exhaustive search method. The HCP parameters are selected by the encoder using the first picture of each Intra Random-Access Point (IRAP) and signalled once per IRAP. In SPS, projection format, frame packing parameters including number of faces in horizontal and vertical directions and each face's position and orientation are signalled. In PPS, the horizontal and vertical HCP parameters in 6-bit precision are encapsulated. The proposed HCP solution is implemented upon JEM-6.0 and 360Lib-3.0 software. Simulation results are reported using the test conditions specified in the JVET Call-for-Evidence (CfE) document. Compared with the CMP and ACP formats, the proposed HCP format demonstrates average 3.0 dB (up to 3.6 dB) and 0.2dB (up to 0.4 dB) End-to-End WS-PSNR improvement for the luma (Y) component, respectively, and average luma (Y) BD-rate reductions of 11.5% (up to 23.0%) and 0.5% (up to 1.0%), respectively.

Run-length coding (RC) is a simple and yet quite effective technique for bi-level image coding. A problem with the conventional RC which describes an image by alternating runs of white and black pixels is that it only exploits the redundancy within the same scan line. The modified relative address run-length coding (MRC) used in Group III facsimile transmission is more efficient by making use of the correlation between adjacent lines. The paper presents a vector run-length coding (VRC) technique which exploits the spatial redundancy more thoroughly by representing images with vector or black patterns and vector run-lengths. Depending on the coding method for the block patterns, various algorithms have been developed, including single run-length VRC (SVRC), double run-length VRC (DVRC), and block VRC (BVRC). The conventional RC is a special case of BVRC with block size of 1*1. The proposed methods have been applied to the CCITT standard test documents and the best result has been obtained with the BVRC method. With a block dimension of 4*4, it has yielded compression gains higher than the MRC with k=4 by 15.5% and 22.7%, when using a single and multiple run-length codebooks, respectively.>

Presents a new technique for image reconstruction from partially received information for hybrid video coding systems using DCT and motion compensated prediction and interpolation. The technique makes use of the smoothness property of typical video signals by requiring the reconstructed samples be smoothly connected with their adjacent samples, both spatially and temporally. This is fulfilled by minimizing the differences between neighboring pixels in the current as well as adjacent frames. The optimal solution is obtained through three linear transformations. This approach can yield more satisfactory results than the existing algorithms, especially for images with large motions or scene changes.>