Research On Key Technologies In Distributed Source Coding

Traditional image source coding algorithms, such as video coding standardsMPEG-X and H.26X or still image coding standard JPEG2000, extract the statisticalcorrelation at the encoder. Computational complexity of the encoder is usually muchhigher than the decoder. With the development of electronic technology, a number ofemerging applications such as wireless video sensor network and camera array havebeen widely used. These new applications have constrained resources and powerconsumption at the encoder. It’s not very suitable to apply the traditional image sourcecoding algorithm in these applications. This is a new challenge to traditional imagecoding algorithms and system architectures.In recent years, Distributed Source Coding (DSC) is being researched by more andmore scholars due to its low encoding complexity, good performance and better errorresilience robustness. Distributed source coding concept was first proposed by Slepianand Wolf in the1970s for lossless compression. They proved that two independentcorrelated sources can achieve the same compression efficiency compared withcompress jointly. Wyner and Ziv then extended it to lossy compression and got similarresults. Slepian-Wolf and Wyner-Ziv two theories laid the theoretical basis of theDistributed source coding. The distributed source coding extract statistical correlation atthe decoder which transfers the complexity from the encoder to the decoder comparedwith traditional image source coding algorithms. Distributed source coding has nowbecome a new research hotspot in worldwide. In this paper, several key technologies inDistributed Source Coding are researched and appropriate solutions are proposed.The main work and research results obtained include:1. A transform-domain distributed source coding architecture is proposed based onresearch of Turbo-based and LDPC-based system structures. At encoder, DCT transformis performed first. Then, bitplane data of subbands after DCT transform and quantitize isencoded with LDPCA. At decoder, side information reconstructed by key frames is usedfor LDPCA decoding. Finally encoded Wyner-Ziv frame is reconstructed with decodedbitplane data. With DCT transform and quantitize technology, this system can achievebetter performance compared with traditional intra-frame coding method.2. A new Distributed Source Coding scheme without feedback channel is proposedto overcome the shortcomings of the feedback channel. The encoder uses fast motionestimation algorithm to reconstruct the side information, and use the Laplace distribution model to estimate the proper bit rate. Since the feedback channel is removed,the system complexity is greatly reduced. The rate distortion performance loss is controlin0.3dB or less and this scheme is suitable for those scenarios feedback channel doesnot exist or high system delay is unacceptable.3. An improved side information interpolation algorithm is presented. We set thefunctional threshold to ensure the linear continuity of target motion track. For smoothmotion area, bi-direction motion estimation and compensate is adopted to generate theside information. For other area, side information macroblock is chosen from a series ofseveral candidate macroblocks with minimum edge match error. Optimal minimummean-error reconstruction algorithm is also adopted to enhance the reconstructionperformance. Experiments show our algorithm can achieve better PSNR performancewhile decreases the encoding rate.4. Based on the analyses of the hyper-spectral images, a new compressionalgorithm based on DCT transform domain distributed source coding is proposed. Itperforms the bitplane encoding at the encoder with DCT subbands order, while usingthe key frame to reconstruct the side information for LDPC decoding at the decoder.Few pixels are adopted to perform linear prediction at encoder which reduces thecomplexity. Subbands which previously decoded are utilized for iterative linearprediction based on blocks at decoder, and following subbands are decoded withoptimized side information. The experimental results show that the proposed algorithmachieves improved performance over the conventional algorithm, and efficientlyreduces the cost of computation and memory usage at the encoder which facilitates thehardware implementation.5. Channel coding based Slepian-Wolf decoding takes up more than90%of theentire decoding complexity of the Wyner-Ziv system. A high performance CUDAparallel computing of distributed source decoder is proposed. GPU basedimplementation obtains up to200x speedup for LDPCA decoding. Compared totraditional CPU-based implementation, GPU parallel computing optimization achieves aspeedup of10x for QCIF decoding, and speedup of20x for CIF decoding. With highperformance parallel computing technology Distributed Source Coding is becomingavailable for practical application.For the above algorithms, a large number of simulations and experiments areperformed to verify their validity and advantages.