An IBC Reference Block Enhancement Model Based on GAN for Screen Content Video Coding

As a special kind of video coding, screen content coding (SCC) has received widespread attention because of the popularity of online classes and conferences. However, few people use neural networks to improve the compression efficiency of SCC. Intra block copy (IBC) is one of the most important coding tools in SCC, which can save half of the bitrate. Due to the need to copy the content of the reference block, the performance of IBC mode largely depends on the quality of the reference block. In the standard encoding process of Versatile Video Coding (VVC), the IBC reference block is not filtered, and there are still serious compression artifacts. This will result in a decrease in IBC search accuracy and SCC compression efficiency. Inspired by in-loop filtering, we propose an IBC reference blocks enhancement network based on GAN (IREGAN) to filter the reference blocks before IBC estimation, which can improve the quality of IBC reference block and the accuracy of IBC matching. In addition to the generator used for image enhancement, our model also includes a variance-based classifier and a discriminator obtained from adversarial training The classifier can effectively improve the efficiency of the model and the discriminator can improve the robustness of the entire system. Experimental results demon-strate the performance gains of IREGAN with VTM10.0, offering about 6.98% BDBR reduction, 0.71dB BDPSNR gains in average (luminance). SSIM increased by 0.0113 and the number of blocks using IBC mode is increased by 1.42%.