Distributed Compressed Video Sensing Based On Landweber Iteration Reconstruction

With the widely use of video communications technology, research in someresource-constrained field of video communications application wins more and moreattention, such as video surveillance, wireless PC cameras, mobile video calls in wirelesssensor networks with limited computing power, memory capacity, and power consumption.How to transmit video effectively high-quality and stable for these special occasionsbecome the higher goals of researchers. DCVS (Distributed Video Compressed Sensing)gives solutions for such problems, it is a new video codec combined with distributedsource coding characteristics of DVC (Distributed Video Coding) and the theoreticaltechniques of CS (Compressed Sensing). It has lower complexity encoder, suitable forreal-time and non-real-time applications. So it has a broad application prospects.In this paper, we focus on the problems that the decoding quality of DCVS is not ideal,and no suitable quantitative methods for measurements, and the codec structure is notoptimal.Then we do in-depth research on measuring methods, quantization algorithms, andLandweber iteration image reconstruction algorithms for DCVS, my research contents andinnovation as follows:1) As exploiting a fixed measurement rate for each block when compressed videosening, resulting in inaccuracy of image reconstruction and obvious blocking artifact.Consequently block-level dynamic measurements rate allocation algorithms in the pixeldomain and wavelet domain are proposed repectively. We use the commonly method ofedge detection to classify the image blocks in the pixel domain. According to the differentcategories to allocate different number of measurements, in the wavelet domain, exploitingthe improved weighted balanced measurements allocation algorithm based ondecomposition level, to allocate the limited number of measurements to each block in terms of its energy proportion. Experiments shows that, two algorithms applied tosmoothed projected Landweber(SPL) image reconstruction have both made a goodreconstruction effect.2) Since measurements quantization is always ignored for compressed video sensing,we propose a new measurements quantization algorithm based on multihypothesisprediction for compressed video sensing. The main idea of this algorithm is that classifythe blocks into “slow” block and “severe” block in terms of the correlation between themeasurements of current block and its neighboring block. Measurements of “slow” blockdirectly get differential quantization coding, for another we abtain the linear combinationcoefficients of measurements belonging the “severe” block by multihypothesis prediction.After that we get the optimal prediction measurements of current block, then the residualsbetween actual measurements and prediction, combined with coefficients are transmittedinto channels after to be quantified and coded. Experiments show that the proposedalgorithm has good codec performance.3) As the utilization of the side information are insufficient for DCVS, resulting in theperformance of video reconstruction is not optimal, then a new Landweber iterationreconstruction algorithm based on dictionary learning denoising is proposed. We regard thetrained dictionary as SI which is exploited for image denoising during the process ofLandweber iterations at decoder. Keeping the encoder consistent, compared withstate-of-the-art DCVS codec, the decoding efficency is higher, especially in the case of lowmeasurement rate. The video frame can be also reconstructed with an acceptableperformance.