Image Restoration Based on CNN-Shallow Residual Learning
TL;DR
A Shallow Residual Network (SRNet) based on convolutional neural network is proposed to focus on the high-frequency detail compensation lost in the compression process, and the ResNet-50 starting layer pre-trained on ImageNet is used for multi-scale feature collaborative extraction.
Abstract
With the development of cloud gaming and esports live streaming, high-fidelity transmission of critical game footage under low bandwidth has become a significant challenge. Traditional JPEG compression is prone to distortion, making it difficult to meet the image quality and fairness requirements of esports scenarios. To this end, this paper proposes a Shallow Residual Network (SRNet) based on convolutional neural network. The residual learning mechanism is introduced to focus on the high-frequency detail compensation lost in the compression process, and the ResNet-50 starting layer pre-trained on ImageNet is used for multi-scale feature collaborative extraction. Three kinds of special datasets including UI sensitivity, special effects detail and map texture were constructed based on the “League of Heroes”. Under the same bit rate condition, the average PSNR of SRNet on UI subset reached 32.61 dB, which was 11.56 dB higher than that of JPEG, and the SSIM increased to 0.8956. The residual analysis shows that the method effectively suppresses the ringing effect in the edge region, and the distortion energy is reduced by more than 70%.