Side Information Fusion And Reconstruction For Distributed Video Coding

Distributed Video Coding (DVC) is a new concept based on Slepian-Wolf lossless and Wyner-Ziv lossy compression theories, where several sources with the same distribution are independently encoded and jointly decoded. Different from conventional video coding technologies: conventional video coding technologies exploit correlations at the encoder side, thus making the encoding complexity 5~10 higher than decoding complexity. Conventional video coding applies well to video broadcasting, video on demand (VOD) and video disk storage, etc; while DVC takes advantage in low encoder complexity, low power consumption and strong error resilience. Thus, DVC has its application in wireless video terminator such as wireless video monitoring system and video sensor network which have limited computation ability, storage capability and power consumption.Distributed Multi-view Video Coding (DMVC) can be regarded as a extension of DVC. Different from DVC which has only temporal side information, there exist side information generated from inter-view direction as well in DMVC. Although the quality of temporal side information is better than that of inter-view side information, from local aspect, inter-view side information is better than temporal one where the motion is high. The task of side information fusion is to find the better side information in the same area and select it as the final side information.For mono view DVC, we proposed a bit plane prediction based method for Wyner-Ziv reconstruction method. Common DVC lose information during the side information interpolation at the decoder. In this paper, we extract features from the adjacent frames which are used to interpolate the side information so as to predict the next bit of the reconstructed frame. The predicted bit is used to modify the probability density function (PDF) of reconstruction. The experimental results show that our method achieves higher PSNR and better subjective effects when fewer bit planes are transmitted.In addition, for DMVC, we proposed Conditional Random Fields (CRF) based fusion method used to combine side information from both temporal and inter-view direction. Common side information fusion method find abrupt motion vectors in temporal side information and replace the block area with inter-view side information. These methods based on block search at decoder and will result in block artifacts. Though feature function definition, we learned the weights among previously decoded Wyner-Ziv frame, current temporal and inter-view side information, thus inferring the current labels and generate the final side information. Our method introduce the spatial correlation into the CRF model which helps recognize abrupt motion areas effectively and replace these area with inter-view side information. Experimental results show that the method is better than current fusion method in rate-distortion performance.