A Hierarchical Two-Stage Neural-Classifier for Mode Decision of H. 264/AVC Stereo Video Encoding

This paper presents a design of H.264-based stereo video encoder. For our system, the currently developed JMVM platform is adopted, by which the encoding of the left-view channel is purely based on the predictions from the temporal domain, while for the right-view channel, combined predictions from the temporal/ disparity domains are exploited, with a hierarchical two-stage neural classifier for fast mode decision. The first-stage classifier determines promising candidates for variable block partition of each MB, while the second-stage classifier aims to choose the most probable prediction sources among the forward/backward motion and disparity frames. The input features for both stages of neural classifiers come from simple calculations on difference images between the currently coded frame and its reference frames. Experiment results indicate that the proposed algorithm can achieve up to 84% of time savings with nearly ignorable quality degradation and acceptable bit-rate increase (less than 7%).