A Convolutional Neural Network-Based Complexity Reduction Scheme in 3D-HEVC

Abstract

View synthesis optimization (VSO) is one of the core techniques for depth map coding in three dimensional high efficiency video coding (3D-HEVC). It improves the quality for synthesized views, while it also introduces heavy computational complexity caused by the calculation of synthesized view distortion change (SVDC) in practice. To reduce the complexity, this paper proposes a convolutional neural network-based VSO scheme in 3D-HEVC. First, the potential factors that may relate to the encoding complexity are explored. Then, based on this, a convolutional neural network (CNN) is embedded into the 3D-HEVC reference software HTM16.0 to predict the depth of coding units (CUs). The complexity of SVDC can be drastically reduced by avoiding the brute-force search for VSO in depth 0 and depth 1. Finally, for depth 2 and depth 3, the zero distortion area (ZDA) is determined based on texture smoothness and the SVDC calculation for that area is skipped. The experimental results show that the proposed scheme can reduce 76.7% encoding time without any significant loss for the 3D video quality.