Semantic-driven bitrate optimization algorithm

• Published in: 中国邮电高校学报（英文版） Vol. 31; no. 6; pp. 76 - 87
• Main Authors: Chen Feng, Huang Faren, Chen Pingping, Chen Yanying
• Format: Journal Article
• Language: Chinese; English
• Published: School of Advanced Manufacturing,Fuzhou University,Jinjiang 362200,China 01.12.2024; College of Physics and Information Engineering,Fuzhou University,Fuzhou 350000,China%School of Advanced Manufacturing,Fuzhou University,Jinjiang 362200,China%College of Physics and Information Engineering,Fuzhou University,Fuzhou 350000,China
• Subjects: object detection; video semantics; bitrate control; machine vision
• ISSN: 1005-8885
• DOI: 10.19682/j.cnki.1005-8885.2024.1024

Traditional rate optimization algorithms typically use peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as optimization objectives to improve video quality while reducing bitrate.However,in machine vision applications,machines are usually not concerned with video quality but rather the semantic information conveyed by the video.In this paper,focusing on the application scenario of connected vehicles,a bitrate control algorithm for the 5 th generation mobile communication technology (5 G) roadside perception cameras in the context of connected vehicles is studied.Specifically,a semantic-driven rate optimization algorithm is proposed.This algorithm introduces a video semantic task model and adaptively allocates bitrate by dividing video frames into multiple independent encoding regions based on video content features.It effectively controls the bitrate while maximizing the semantic task.Experimental results show that compared with the high efficiency video coding (HEVC) test model HM16.23 encoder using the bidirectional mainstream largest coding unit (LCU)-layer model bitrate control algorithm,the semantic-driven bitrate optimization algorithm (SDBOA) proposed in this article saves an average of 10.1％ of the bitrate.Compared with the HMLCU1 algorithm,the target detection accuracy is improved by 3.54％ on average.Compared with the HMLCU2 algorithm,the target detection accuracy is improved by 7.54％ on average.SDBOA is more in line with the current video processing scenarios in mainstream semantic analysis tasks.