Deep Learning-Based CTU Splitting and Dynamic GOP Adjusting for Semantic Aware Video Compression

The rapid development of the internet has led to a surge in demand for video traffic, presenting new challenges for the storage and transmission of video data. In this paper, considering the semantic information of video frames, semantic aware-based algorithms for coding tree unit (CTU) and group of picture (GOP) have been designed for video encoding. The proposed algorithm utilizes deep learning to forecast the outcomes of CTU splitting in video coding via trained models. Furthermore, a semantic aware GOP spliting approach is developed. Experimental results confirm that the proposed algorithms can improve the performance of video compression. The proposed CTU spliting algorithm saves a significant amount of intra-frame encoding time, and the dynamic GOP spliting algorithm achieves quality improvement at the same bit rate.