Dilated Deep Residual Network for Post-processing in TPG Based Image Coding

Lossy image compression algorithms like JPEG usually introduce visually annoying artifacts on decoded images, such as blocking artifacts, blurring and ringing effects. The tiny portable graphics (TPG) based image/video compression technique is proposed to improve JPEG compression performance. However, the lossy compression artifacts cannot be fully removed, especially at low coding bit-rates. Recently, some shallow convolutional neural network (CNN) models have been proposed as post-processing techniques to reduce compression artifacts. Learning from the fact that deep CNNs have shown extraordinary ability in high-level vision problems, we propose to investigate how a deeper CNN can further enhance the quality of decoded images. Specifically, we adopt a network with 16 residual blocks. In order to increase the receptive field, we change the first convolution layer in the first five residual blocks to dilated convolution with size 2. The primary experimental results show that the proposed model can outperform existing CNN based post-processing methods.