CU Split Method Based on Adaptive CNN and Gradient Matrix for VVC 3D Video Depth Map

Abstract

In the digital age, people’s demand for 3D videos is becoming increasingly strong. Nowadays, the 3D-HEVC video coding standard is far from meeting people’s needs. Compared to HEVC, Versatile Video Coding (VVC) exhibits better encoding performance. Depth maps are a critical part of 3D video, but current research on VVC has only been limited to texture videos. Accordingly, to diminish the computational complicacy of depth map intra coding block division in VVC 3D video, a fast approach for VVC depth map coding based on texture characteristics and deep learning is presented in this paper. We first use gradient matrix to classify CUs into simple CUs, fuzzy CUs, and complex CUs. Simple CUs can terminate their partitioning process in advance, while fuzzy CUs use the original encoder algorithm. For complex CUs, we designed two adaptive CCNs that can serve multiple sizes of CUs. The first CNN model is used to determine whether a square CU performs quadtree partitioning or multi type tree partitioning. The second CNN model is used to determine whether a CU that definitely performs multi type tree partitioning performs horizontal or vertical tree partitioning. It is clear that the second model is complementary to the first. The experimental results indicate that this scheme can achieve an average reduction of 45.35% in coding time, while BDBR only increases by 0.23%, which is superior to existing technologies in reducing encoding complexity and ensuring encoding quality.

Keywords

• Depth map
• early termination
• VVC 3D video
• CNN