||Study on Low Bit-Rate Distributed Compressive Video Sensing
Kurihara, Shin栗原, 信
47 , 2017-09-20 , 新潟大学
学位の種類: 博士（工学）. 報告番号: 甲第4378号. 学位記番号: 新大院博（工）甲第470号. 学位授与年月日: 平成29年9月20日
The thesis deals with a new architecture of video codec that is based on compressive sensing (CS) and distributed video coding (DVC). One of the critical problems is the issue of high bit rates of the distributed compressive video sensing (DCVS) studied in the video coding field in recent years. In this study, all frames are simply sampled by the CS measurement at the observation phase, then they are encoded by DVC. It is a remarkable feature different from the other existing approaches. At the encoder, CS-observed key frames are encoded by H.264/AVC intra prediction. On the other hand, compressively-sensed non-key frames are encoded by means of the inter frame difference compression and Wyner-Ziv (WZ) encoding. Details are described at 3 chapter about a decoder. The transmission data amount to the decoder is reduced very much compared with the conventional DCVS. The objective and subjective quality of decoded video is evaluated in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index as well as visual inspections. In Chapter 1, the background of this study and the aim are described. Recent years the Internet of things (IOT) is rapidly in progress. The development of video applications in wireless cameras and vehicle-mounted cameras are active. In those applications, a few critical problems are the cost of a front-end terminal and the power consumption in battery-driven devices. In order to solve these problems, combinations of CS and DVC are tried actively. This chapter overviews some of current studies to address the problems. In the next, DCVS is described at the observation phase, in which every frame is equally subjected to CS at the observation phase. DCVS offers a very simple encoder with a low burden, while one of serious problems is the transmission data volume that is very large between encoder and decoder. It is well-known that the communication data amount very influences the power consumption. Therefore, in this study, we propose a novel architecture in which every frame is compressively sampled with CS, before it is subjected to either predictive coding or distributed coding.The purpose of this study is to reduce the data volume in transmission by this method.
In Chapter 2, basic theories and technology about CS, DVC and DCVS are described by some expressions and figures. At first, the CS theory and the gradient projection for a sparse reconstruction (GPSR) are described. Secondly, the theory of DVC and intra prediction of H.264/AVC are overviewed. Finally, DCVS is discussed for a preparation of Chapter 3 and later. In Chapter 3, the architecture of the proposed codec system of the distributed video coding based on CS and intra prediction of H.264/AVC is illustrated. It is shared with a key frame and a non-key frame and is explained in detail. At the encoder, CS-observed key frames are encoded by H.264/AVC intra prediction. At the decoder, the key frames are decoded by intra predictive decoding and a sparse reconstruction. The result shows that the spatial correlation among CS-observed frames is still well preserved to a significant extent before and after CS operation. On the other hand, compressively-sampled non-key frames are encoded by inter frame difference compression and Wyner-Ziv (WZ) encoding to exploit the inter-frame correlation. At the decoder, motion estimation/motion compensation (ME/MC)-based interpolation is applied to key frame pictures to obtain the estimates of inter prediction frames. These predictive frame estimates are refined by Golay decoding with error-correcting data delivered from the encoder. Non-key frames are then obtained by their sparse reconstructions. Finally, a Gaussian lowpass filter and unsharp masking (USM) are applied to every frame to suppress the block artifacts created by the CS reconstruction. The experimental results are shown in Chapter 4, where the proposed and DCVS systems are compared on a key frame, a non-key frame, and a video data unit in a group of pictures. Performance measures are the transmission bit rate between the encoder and decoder, PSNR, SSIM and visual inspections. By these performance evaluations, the proposed method is verified to be successful in very-low bit rate transmission of video sequences. In Chapter 5, conclusions of the thesis and some future works are given.