Neural Video Compression with Re-Parametrisation Scene Content-Adaptive Network

To further improve the video compression performance, a neural network based content-adaptive video compression method is proposed. In current neural video compression methods, the compression quality of I-Frames has a significant impact on the final video compression quality. The compressed I-Frames and P-Frames still have a lot of redundant information. To reduce the redundant information present in compressed I/P frames, we propose a compression scheme with a content-adaptive enhancement for I-Frames and P-Frames, respectively, to effectively improve video compression efficiency. In this paper, three compression quality enhancement modules are designed to compress I and P frames more efficiently. These include the Scene Adaptive Intraframe Image Compression (SAIC) module, which optimizes the encoder parameters for all I-Frame content in a single scene, thus greatly improving the compression quality of I-Frames. The Re-parameterized Scene Adaptive I-Frame Image Quality Enhancement (RSA-QE) module. The Re-parameterized method allows the transmission of the degraded lightweight model, thus improving the video quality as much as possible with a small additional model bitstream cost. Finally, a Scene Adaptive P-Frame Image Compression (SAPC) module is constructed to further enhance the P-Frame compression efficiency. The experimental results demonstrate that the proposed method is competitive with state-of-the-art.