基于通道注意力的协作智能特征压缩算法  

A Collaborative Intelligent Feature Compression Algorithm Based on ChannelAttention

作　　者：李浩 何小海[1] 陈洪刚[1] 杨红[1] 熊淑华[1] LI Hao;HE Xiaohai;CHEN Honggang;YANG Hong;XIONG Shuhua(College of Electronic Information,Sichuan University,Chengdu Sichuan 610065,China)

机构地区：[1]四川大学电子信息学院,四川成都610065

出　　处：《通信技术》2024年第3期236-243,共8页Communications Technology

基　　金：国家自然科学基金(62271336,62211530110);TCL科技创新基金;四川省科技项目基金(24GJHZ0381)。

摘　　要：由于越来越多的视频开始使用机器视觉进行任务分析,以往面向人眼的视频压缩方法难以保障视频目标检测等任务性能。对此,提出一种基于通道注意力的协作智能特征压缩算法。首先,选择分裂点特征,将其作为特征压缩网络的输入;其次,使用通道注意力对特征进行通道冗余去除,使用卷积进行空间冗余去除;最后,运用广义除法归一化减少噪声,以确保重建特征的还原度。实验结果表明,该特征压缩算法可以有效缓解通信带宽压力,高质量重建特征,保障目标检测任务性能。Due to the increasing use of machine vision for task analysis in videos,previous video compression methods designed for human vision struggle to ensure task performance such as video object detection.In response to this challenge,a collaborative intelligent feature compression algorithm based on channel attention is proposed.First,this paper selects a splitting point feature and uses it as input to the feature compression network.Then,it employs channel attention to eliminate channel redundancy and uses convolution to eliminate spatial redundancy.Finally,it uses generalized division normalization to mitigate noise and to ensure the restorative quality of the reconstructed features.Experimental results indicate that the feature compression algorithm can effectively alleviate communication bandwidth pressure,reconstruct features with high quality,and ensure the performance of the object detection task.

关 键 词：机器视觉 视频压缩 目标检测 协作智能 特征压缩 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]