CNN-Based Inter Prediction Refinement for AVS3

The third generation of audio video coding standard, also known as AVS3, is the latest audio video coding standard developed by China AVS working group. With the newly adopted coding tools, AVS3 has achieved 20% coding gain compared to HEVC. To further improve the coding performance of AVS3, a CNN-based inter prediction refinement algorithm is proposed in this paper. A fully convolutional neural network structure is adopted to enhance CUs with different shapes. In order to ensure the network robust for various CU shapes, a progressive training scheme is applied. Experimental results demonstrate that the proposed method achieves on average 1.88% and 1.17% BD-rate gain under LD and RA configuration following the common test condition of AVS3.