Image coding based on Selective Super-Resolution Network

This paper presents an image compression method using deep learning based super-resolution. Different from traditional image compression methods, we propose to select the super resolution network to enhance the perceptual quality. In the proposed method, the network selection map is generated by utilizing the semantic segmentation, image classification, and encoding distortion. Compared with traditional scalable image coding methods, more texture area are visually pleased even in the severe low bitrate encoding. For this reason, our method has the potential to result better compression ratio in terms of subjective image quality. Experimental results demonstrate the effectiveness of the proposed method. When compared with high efficiency video coding (HEVC), our method achieve the average of 43.42% bitrate saving at the same perceptual image quality.