A perceptual quantization strategy for HEVC based on a convolutional neural network trained on natural images

摘要

Fast prediction models of local distortion visibility and local quality can potentially make modern spatiotem-porally adaptive coding schemes feasible for real-time applications. In this paper, a fast convolutional-neural-network based quantization strategy for HEVC is proposed. Local artifact visibility is predicted via a network trained on data derived from our improved contrast gain control model. The contrast gain control model was trained on our recent database of local distortion visibility in natural scenes [Alam et al. JOV 2014]. Furthermore, a structural facilitation model was proposed to capture effects of recognizable structures on distortion visibility via the contrast gain control model. Our results provide on average 11% improvements in compression efficiency for spatial luma channel of HEVC while requiring almost one hundredth of the computational time of an equivalent gain control model. Our work opens the doors for similar techniques which may work for different forthcoming compression standards.