The traditional image compressors, e.g., BPG and H.266, have achieved great image and video compression quality. Recently, Convolutional Neural Network has been used widely in image compression. We proposed an attention-based convolutional neural network for low bit-rate compression to post-process the output of traditional image compression decoder. Across the experimental results on validation sets, the post-processing module trained by MAE and MS-SSIM losses yields the highest PSNR of 32.10 on average at the bit-rate of 0.15.

Attention Based Image Compression Post-Processing Convolutional Neural Network