A Deep Learning Based Perceptual Bit Allocation Scheme on Conversational Videos for HEVC \lambda -Domain Rate Control

The newest \(\lambda \)-domain rate control in High efficiency video coding (HEVC) adaptively allocates bit per pixel (bpp) without considering the visual importance. A perceptual bit allocation scheme based on deep learning for conversational videos is proposed. Firstly, a multitask cascaded convolutional network is employed to detected facial region in the encoding videos. Then, instead of the predefined bit ratio for frame level bit allocation, an adaptively bit rate ratio, mainly according to variation of inter frames, is used to allocate bit rate more reasonable for every frame. Finally, the quality parameters (QP), belonging to the facial regions, will be clipped in a smaller interval to get a better visual quality. The experimental results show that the quality of facial region is improved significantly with a good rate control performance.