Research On Distributed Video Processing Algorithms Based On Compressive Sensing

In real-time video sensing of field ecological environment,wireless multimedia sensor networks resources are limited.Distributed compressive video sensing,which combines distributed video coding and compressive sensing,provides a new insight into this application.This dissertation focuses on distributed compressive video sensing.Aiming to improving the reconstruction quality,reconstruction efficiency,and user experience,this disseratation explores high-quality,efficient,and robust algorithms and provides theoretical support for the practical application of distributed compressive video sensing.Detailed contributions and innovations of this dissertation are as follows:1.Gradient projection for sparse reconstruction(GPSR)-based fast distributed compressive video sensing framework is vulnerable to inaccurate image group selection algorithms,which can further cause video flicker.To address this problem,a perceptual hash-based adaptive selection algorithm of groups of pictures is proposed.Specifically,the algorithm first proposes an adaptive selection algorithm of groups of pictures which utilizes the perceptual hash algorithm to measure the similarity between adjacent video frames,and then carries out the adaptive selection.On this basis,an adaptive frame sampling rate allocation algorithm,which first establishes a constraint optimization model based on the characteristics of distributed compressive video sensing and then allocates frame sampling rates by solving the constraint optimization problem,is proposed to further improve the overall reconstruction quality.The experimental results demonstrate that the proposed algorithm can efficiently alleviate video flicker.2.A hybrid multi-hypothesis prediction reconstruction algorithm based on measurement reorganization is proposed,which can prevent quality deterioration caused by the high-probability failure of theoretical conditions of the multi-hypothesis prediction algorithm under the condition of low sampling rate.Specifically,the algorithm first proposes a multi-hypothesis prediction reconstruction algorithm based on measurement reorganization which reorganizes the measurements of the non-key frame and the side information to improve the probability of satisfying the theoretical conditions of multi-hypothesis prediction algorithm and then carries out the multi-hypothesis prediction and the residual reconstruction to obtain the initial reconstruction.On this basis,a hybrid multi-hypothesis prediction reconstruction algorithm based on resampling,which carries out both the traditional multi-hypothesis prediction reconstruction and the measurement reorganization-based multi-hypothesis prediction reconstruction to obtain multiple initial reconstructions and then adaptively mixes multiple initial reconstructions by measuring the similarity between the original measurements of the non-key frame and the resampled measurement of the initial reconstruction to obtain the final reconstruction,is proposed to prevent the negative effect of inaccurate side information.The experimental results demonstrate that the proposed algorithm can improve reconstruction quality when having a low sampling rate.3.The multi-hypothesis prediction algorithm holds the common assumption that hypothesis that are dissimilar from the target block are given lower weights than that are more similar,which can easily neglect detailed information.To address this problem,a joint sparse model-based multi-hypothesis prediction reconstruction algorithm is proposed.The algorithm proposes a new assumption that the multi-hypothesis prediction of the target block consists of the common information which is provided by the most similar hypothesis and the detailed information which is provided by other hypotheses.Based on this assumption,a hypothesis generation scheme which divides the initial set of hypotheses into two subsets,one for the common information and the other for the detailed information.On this basis,a Tikhonov regularization matrix based on Euclidean distance is proposed for the detailed information.The experimental results demonstrate that the proposed algorithm can efficiently enhance detailed information.4.Robustness of the multi-hypothesis prediction algorithm is vulnerable to the accuracy of hypotheses,the number of hypotheses and the accuracy of regularization.To address this problem,an iterative reweighted multi-hypothesis prediction reconstruction algorithm is proposed.Specifically,the algorithm proposes a multi-hypothesis prediction reconstruction algorithm based on reweighted Tikhonov regularization which comprehensively considers the influence of the above three factors on the performance of the multi-hypothesis prediction algorithm by measuring the influence of each hypothesis to enhance the robustness of the multi-hypothesis prediction algorithm.On this basis,the algorithm proposes a Bhattacharyya coefficient-based stopping criterion for the recovery of non-key frames to avoid over-iteration and further enhance the robustness which utilizes the Bhattacharyya coefficient to measure the similarity between the residuals of the non-key frame and that of the adjacent key frame,and constructs the stopping criterion.The experimental results demonstrate that the proposed algorithm can efficiently enhance the robustness of the multi-hypothesis prediction algorithm.5.Inferior reconstruction efficiency make the existing distributed compressive video sensing frameworks hard to meet the real-time application requirements.To address this problem,a joint sampling-reconstruction network based on convolutional neural networks is proposed.Specifically,the proposed network utilizes deep learning technology to build an end-to-end,learnable convolutional neural network.On the basis of independent encoding and joint decoding,the proposed network realizes ‘joint coding and joint decoding' which makes full use of inter-frame correlation and intra-frame correlation at both encoder and decoder.The experimental results demonstrate that the proposed network has high reconstruction quality and efficiency.