Learning a virtual codec based on deep convolutional neural network to compress image

Abstract

In this paper, we propose a standard-compliant image compression framework based on image representation network (IRN) and post-processing neural network (PNN), which are trained by learning a virtual codec network (VCN). Firstly, we use a mixed-resolution image coding considering different types of distortions caused by image compression with different quality factors. Secondly, the VCN is introduced to learn a differentiable soft-projection from the represented image to the post-processed image to resolve the non-differentiable problem of hard quantization. Thirdly, the PNN is used to greatly enhance the quality of decoded images, since standard codecs always result in visually unpleasant blocking artifacts and ringing artifacts. Finally, our framework is trained in an end-to-end manner, whose convolutional kernels of the IRN, PNN and VCN are initialized by pre-training an auto-encoder network. Experimental results verify that our method has higher coding efficiency than the newest image representation-based compression method and many post-processing approaches.