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

摘要

Although deep convolutional neural network has been proved to efficientlyeliminate coding artifacts caused by the coarse quantization of traditionalcodec, it's difficult to train any neural network in front of the encoder forgradient's back-propagation. In this paper, we propose an end-to-end imagecompression framework based on convolutional neural network to resolve theproblem of non-differentiability of the quantization function in the standardcodec. First, the feature description neural network is used to get a validdescription in the low-dimension space with respect to the ground-truth imageso that the amount of image data is greatly reduced for storage ortransmission. After image's valid description, standard image codec such asJPEG is leveraged to further compress image, which leads to image's greatdistortion and compression artifacts, especially blocking artifacts, detailmissing, blurring, and ringing artifacts. Then, we use a post-processing neuralnetwork to remove these artifacts. Due to the challenge of directly learning anon-linear function for a standard codec based on convolutional neural network,we propose to learn a virtual codec neural network to approximate theprojection from the valid description image to the post-processed compressedimage, so that the gradient could be efficiently back-propagated from thepost-processing neural network to the feature description neural network duringtraining. Meanwhile, an advanced learning algorithm is proposed to train ourdeep neural networks for compression. Obviously, the priority of the proposedmethod is compatible with standard existing codecs and our learning strategycan be easily extended into these codecs based on convolutional neural network.Experimental results have demonstrated the advances of the proposed method ascompared to several state-of-the-art approaches, especially at very lowbit-rate.