Cost Sensitive Learning Based HEVC Screen Content Intra Coding for Mobile Devices

Yuanyuan Xu, Kun Zhu

Abstract: The distinguished features of screen contents lead to the development of the Screen Content Coding (SCC) extension of High Efficiency Video Coding (HEVC). Intra coding of HEVC-SCC has to evaluate traditional HEVC intra coding modes and additional modes for screen contents, which poses a challenge for coding such a content on mobile devices. Furthermore, the heterogeneous mobile devices have varying complexity requirements. In this paper, a flexible screen content intra coding scheme is proposed, which can trade between encoding complexity and rate-distortion performance degradation via cost sensitive learning. Taking the cost of different decisions into account, binary classifiers are designed to determine whether a particular kind of coding modes will be evaluated. By adopting the cost-proportionate weighted sampling, the proposed scheme can learn a flexible coding policy offline via reinforcement learning. Experimental results show the effectiveness of the proposed scheme.