Multiple description coding network based on semantic segmentation

Considering semantic information in the image compression can prominently improve the quality of synthesized image. In this paper, we propose a multiple description coding network based on semantic segmentation. In the proposed scheme, the semantic segmentation map of input image is encoded as side information to improve the coding efficiency. Firstly, multiple description feature generator network is used to produce multiple description information. Secondly, the produced multiple description information and the semantic segmentation map are fed into the semantic segmentation encoder network to obtain encoded information. Thirdly, we propose side decoder networks and central decoder network, which are used to decode the image. In the proposed architecture, the semantic information is auxiliary information, which is used to compensate the difference between the input image and generated image. After testing the two datasets, it can be seen that when the bit rate is greater than 1BPP, the PSNR can exceed 40. Therefore, the proposed method is feasible.