Artificial Intelligence Based Adaptive GOP Size Selection for Effective Wyner-Ziv Video Coding

Wyner-Ziv video coding (WZVC) has been gaining many attentions in recent decades due to its low computational complexity and error resiliency benefits, notably when compared to traditional video coding standards such as H.264/AVC or High Efficiency Video Coding (HEVC) standards. In a Wyner-Ziv video coding scheme, the compression efficiency can be controlled by the length of the group of pictures (GOP) which typically consists of the two key and several WZ frames. However, the current Wyner-Ziv video coding solutions usually employ a fixed GOP size or simple adaptive GOP size mechanisms, which depend on some heuristic features extracted from video content. To address the limitation of the current GOP size adaptation solutions, we propose in this paper a novel Artificial Intelligence based GOP size adaptation mechanism and integrate it into the most advanced transform domain Wyner-Ziv video coding (TDWZ) architecture. In the proposed GOP size adaptation mechanism, the proper GOP size is learnt from the correlation between video features and the optimal compression performance. The power of machine learning techniques is used to select the most suitable video features and the model of GOP size and compression performance correlation. Experimental results shown that, using the obtained GOP size adaptation mechanism, the TDWZ codec achieved a better compression performance when compared to relevant benchmarks.