Semantic Map Based Image Compression via Conditional Generative Adversarial Network

Zhensong Wei, Zeyi Liao, Huihui Bai, Yao Zhao

Abstract: Recently, deep learning methods have been applied for image compression and achieved promising results. For lossy image compression at low bit rate, the traditional compression algorithms usually introduce undesired compression artifacts, such as blocking and blurry effects. In this paper, we propose a novel semantic map based image compression framework (SMIC), restoring visually pleasing images at significantly low bit rate. At the encoder, a semantic segmentation network (SS-Net) is designed to generate a semantic map, which is encoded as the first part of the bit stream. Furthermore, a sampled image of the input image is compressed as the second part of bit stream. Then, at the decoder, in order to reconstruct high perceptual quality images, we design an image reconstruction network (Rec-Net) conditioned on the sampled image and corresponding semantic map. Experimental results demonstrate that the proposed framework can reconstruct more perceptually pleasing images at low bit rate.