Research On Entropy Coding For Video And Image Compression

As one of the most important modules in video/image codec, the entropy coding plays an indispensable role in compressing the sources symbols and organizing the bit-stream. The entropy coding first utilizes the context modeling scheme to exploit the statistical redundancy, and then utilizes the entropy coding engines i.e. variable length coding or arithmetic coding to remove the statistical redundancy and generate the compact bit-stream. The entropy coding usually uses the previously coded symbols to estimate the probability of the current symbol, thus they can efficiently remove the statistical redundancies. However, the coding dependencies make it difficult to improve the throughput of video codec. Therefore, how to banlance the throughput and coding efficiency has been the hot topic in video coding fields. With the explosion of the high quality video, the future video coding standards have to process the video with massive data, so it is still the main target to improve the coding efficiency when establishing the future video coding standards.To overcome the restrictions of external conditions, the compressive sensing(CS) has been applied in the image/video acquisition in the emerging multimedia applications. The obtained measurements are different from the pixels in nature when applying CS into the acquisition of video/image. More specificially, the measurements are obtained via a linear projection into a lower-dimensional subspace chosen at random. Thus each measurement contains the information from all pixels in one image and one measurement is independent of others. So it is unsuitable to apply the conventional video/image coding tools to compress the measurements. Therefore, it is desirable to develop coding tools for compressing the CS measurements.So, in this context described above, this dissertation focuses on entropy coding modules in the video coding standards including H.264/AVC, HEVC and AVS2 as well as the image compression based on compressive sensing.The main contributions are composed of the following four works:First, since the throughput is not fully taken into account in CABAC within H.264/AVC, CABAC has been one of the most time-consuming modules in the decoder. To improve the throughput of CABAC, a hierarchical dependency context model(HDCM) is firstly proposed to exploit the statistical correlations of DCT coefficients, in which the number of significant coefficients in a transform block and the scanned position are used to capture the magnitude varying tendency of transform coefficients. Then a new binary arithmetic coding using HDCM(HDCMBAC) is proposed. HDCMBAC associates HDCM with binary arithmetic coding to code the syntax elements for a transform block, which consist of the number of significant coefficients, significant flag and level information. Experimental results demonstrate that HDCMBAC can achieve similar coding performance as CABAC at low and high QPs. Meanwhile, the context dependency in the context modeling scheme can be reduced as much as possible.Second, to further improve the coding efficiency of entropy coding module and prepare to establish the next generation video coding standard succeeding HEVC, an enhanced entropy coding scheme is proposed based on HEVC, which includes an improved context modeling scheme for transform coefficient levels and an binary arithmetic coding(BAC) engine with low memory requirement. In the improved context modeling scheme for transform coefficient levels, the contexts of significant_coeff_flag consist of the number of the significant transform coefficient levels in a local template and its position. To limit the total number of context models, TBs(TB: transform block) are split into different regions based on the coefficient positions. The same region in different TBs shares the same context model set. The context model index of coeff_abs_greater1_flag is determined by the number of transform coefficient levels in a local template with absolute magnitude equal to 1 and larger than 1. Moreover, TBs are also split into different regions to incorporate the coefficient position in luma in its context model selection. The context model index for coeff_abs_greater2_flag is determined by the number of transform coefficient levels in a local template with absolute magnitude larger than 1 and larger than 2. In the BAC engine with low memory requirement, the probability is estimated based on a multi-parameter probability update mechanism. Moreover, the multiplication with low bit capacities is used in the coding interval subdivision to substitute the large look-up table to reduce its memory consumption. Experimental results demonstrate that the proposed two techniques can significantly improve the coding efficiency of the entropy coding module.Third, there exist strong sequential dependencies in the entropy coding scheme within AVS2, which severely limit the throughput improvement of AVS2. These sequential dependencies mainly stem from normalization process and the bypass bins coding process in the arithmetic coding engine as well as the transform coefficients coding process. So, a fast and standard-complicant normalization is first proposed, which can simplify the arithmetic coding engine. Then, a fast coding process for bypass bins is proposed, in which only addition and shift operations are required for coding the bypass bins. Finally, the context modeling scheme for transform coefficients is modified to reduce the coding dependencies as much as possible. Experimental results demonstrate that the above three techniques can significantly improve the throughput of the entropy coding scheme in AVS2 by keeping the similar coding performance.Fourth, Differential pulse-code modulation(DPCM) is recently coupled with uniform scalar quantization(SQ) to improve the rate-distortion performance for the block-based quantized compressive sensing of images. However, the entropy coding is still required to convert the quantization index of CS measurements into bitstream. Therefore, an arithmetic coding scheme is proposed for the quantization index within DPCM-plus-SQ framework by analyzing their statistics. In the proposed arithmetic coding scheme, the quantization index of the CS measurements is first represented by three syntax elements significant_map, abs_coeff_level_minus1 and sign_flag, which denote the significance(i.e. non-zero), magnitude and sign flag of a quantization index, respectively. These syntax elements are then coded by an adaptive binary arithmetic coding engine M coder to remove the statistical redundancies and generate the bitstream, since M coder can capture the local statistics of these syntax elements. Compared with the zero order entropy of the quantization index and transform coefficient coding in CABAC within H.264/AVC, the proposed arithmetic coding scheme can further improve the coding efficiency.