Deep Image Compression with Residual Learning

Abstract

An end-to-end image compression framework based on deep residual learning is proposed. Three levels of residual learning are adopted to improve the compression quality: (1) the ResNet structure; (2) the deep channel residual learning for quantization; and (3) the global residual learning in full resolution. Residual distribution is commonly a single Gaussian distribution, and relatively easy to be learned by the neural network. Furthermore, an attention model is combined in the proposed framework to compress regions of an image with different bits adaptively. Across the experimental results on Kodak PhotoCD test set, the proposed approach outperforms JPEG and JPEG2000 by PSNR and MS-SSIM at low BPP (bit per pixel). Furthermore, it can produce much better visual quality. Compared to the state-of-the-art deep learning-based codecs, the proposed approach also achieves competitive performance.