Learning a Deep Vector Quantization Network for Image Compression

TL;DR: This paper proposes a DCNN architecture for image compression, where the encoder, quantizer and decoder are jointly learned, and proposes a fully convolutional vector quantization network (VQNet) to quantize the feature vectors of the image representation.

Abstract: Deep convolutional neural network (DCNN) based image codecs, consisting of encoder, quantizer and decoder, have achieved promising image compression results. The major challenge in learning these DCNN models lies in the joint optimization of the encoder, quantizer and decoder, as well as the adaptivity to the input images. In this paper, we proposed a DCNN architecture for image compression, where the encoder, quantizer and decoder are jointly learned. Specifically, a fully convolutional vector quantization network (VQNet) has been proposed to quantize the feature vectors of the image representation, where the representative vectors of the VQNet are jointly optimized with other network parameters through end-to-end training. While most of current DCNN-based methods were only trained once on large-scale datasets, we further perform fine-tuning of the encoder and the codes generated by the VQNet on the input images to further improve the compression performance. Extensive experimental results show that the proposed method achieves state-of-the-art compression results with simple encoder-decoder.