Research On Distributed Video Coding Based On Compressive Sensing

Distributed compressed sensing video using an independent coding and joint decoding scheme reduces the complexity of the encoding side. It is widely used in resource constrained wireless video communications. In the encoder, key frame and non-key frame are encoded respectively to get subsampling values by multiplying each sample matrix. In the decoder, the key frames are recconstructed by the sub-sampling values, but the non-key frames are recconstructed by the sub-sampling values and side information which is generated by the reconstructed key frame. In the distributed compressed sensing video coding framework, the side information accuracy and the reconstruction algorithm directly affect the video reconstruction quality. In order to obtain more precise side information, we improved the method of dictionary’s sample extraction. Furthermore, we improved the iterative projected Landweber thresholding reconstruction algorithm. The main research content and innovation are reflected in the following three aspects:(1) To generate the dictionary testing samples highly relevant to non-key frame, an adaptive weighted side information dictionary generation algorithm was proposed. Motion estimation was performed between two adjacent reconstructed key frames. Then the weighted factors were determined according to the similarity between the above two adjacent motion-estimated key frames and the non-key frame in measured domain. The samples needed in the dictionary training were generated by the adjacent motion estimated key frame and the weighted factors. Therefore, this method takes full advantage of the correlation between two adjacent key frames and non-key frame, and the reconstruction quality of non-key frame has been efficiently improved.(2) In order to improve the precision of side information, we proposed a side information generation algorithm based on multi-hypothesis prediction. First, using two key frames to get the initial side information with bi-directional motion estimation. Secondly, using the initial side nformation and the measured values of non-key frame do the multi-hypothesis prediction, then handle the adjacent two keyframes in the same way. After this, we got three candidate side information. Finally, the block of the final side information was choosed from the three candidate side information according to the correlation between the three candidate side information and the measured value of non-key frame. All of the selected blocks combined into the final side information. Experimental results showed that the reconstructed quality of non-key frames were improved.(3) To overcome the disadvantage of iterative projected-Landweber threshold reconstruction algorithm, the iterative process has been improved. The results of the previous two iterations were added to the next iteration of the iterative process. Then the bias generated in the hard threshold process was effectively reduced and the reconstructed video quality was improved.