Study On Compressive Sensing Multiple Description Coding Scheme

As the main body of the network data, the amount of the image and video signals isincreasing rapidly with the development of information techniques. The current networkcommunication systems mainly have two features. On one hand, the communicationchannels change frequently; on the other hand, the information acquisition systemsbecome more and more miniaturized. These two features make big data encoding andtransmission very challenging. Thus, designing a simple and robust network codingschemes has become an important research topic. Traditional“sampling-compression-protection” based network coding methods have very highcoding complexity, and cannot be used for resource limited network encoders. Recently,as a new theory of sampling, Compressive sensing (CS) provides us a new solution forimage coding. Different from the traditional image encoders that transforms the animage into a sparse domain, the CS-based coding scheme compresses images withuniversal random measurement matrixes, and shifts the task of sparse representation tothe decoders. Such coding scheme is very suitable for current network communicationsystems, which have resource limited encoders and more powerful decoders. In thisthesis, aiming to developing a robust and low-complexity multiple description codingscheme, we propose a CS-based multiple description coding (MDC) method, andinvestigate the quantization, reconstruction algorithm, and measurement issues. Themain contributions of the paper include:(1) A fine granularity CS-based multiple description video coding method has beenproposed. For a low-complexity encoder, a random partial Toeplitz matrix is adopted togenerate fine granularity descriptions. To exploit the spatial and temporal correlations,we propose to use the piecewise autoregressive (PAR) model and total variation (TV)model for effective CS image reconstruction. We also propose to generate adown-sampling based description to facilitate the learning of the varying PAR modelparameters. Compared with traditional unequal error protection (UEP) based MDCschemes, the proposed MDC coding scheme is more robust to resist both erasure and biterrors.(2) A progressive quantization (PQ) for compressive sensing measurements isproposed. In order to maintain the low complexity of the encoder, a scalar uniform quantizer is used. To improve the coding efficiency, we investigate the distribution ofmeasurements, and find that random measurements are not “random”. Instead there areimplicit correlations between these random measurements. To exploit these correlations,a progressive quantization scheme using both scalar quantizer and binned quantizer wasproposed. The dequantization process is recast as an estimation problem to improve thereconstruction of the quantized signal. In addition, we also discussed the practicalimplementation issue and provided a simple solver for the PQ scheme to make thecoding complexity as low as possible. Experimental results show that our PQ schemecan outperform the tradition single-layer quantization (SQ) scheme in both PSNR andvisual quality.(3) We also propose a new image reconstruction algorithm using multiple sparsespaces to further improve the reconstruction quality. Since natural images arenon-stationary, the reconstruction algorithm based on single sparse space usually fails tocharacterize the local image structures. To overcome this drawback, we adopt twocomplementary models to characterize the local sparsity of natural images, andadaptively select the better one according to the local image features. Using Bayesianestimation theory, we propose an estimation method to adaptively select theregularization parameters. In addition, an alternating direction method (ADM) is used toefficiently solve the resulted optimization problem. Experimental results show that theproposed reconstruction algorithm outperforms the method using single sparse space.(4) In order for a low complexity CS-based coding scheme, we propose a hybridsampling scheme using DCT and random measurement matrix. The measurements areencoded by using a Golomb entropy coding scheme, which has low coding complexity.By exploiting the correlation between the two kinds of measurements, the average codelength can be further reduced. Since both DCT and random matrices are incorporatedinto the system, the coding efficiency is improved and a high robust ability to resisterasures is achieved. Experimental results show that the proposed CS-based multiplecoding scheme based on hybrid sampling is more effective than that based on randomsampling.