计算机工程与应用2026，Vol.62Issue(17)：283-294,12.DOI:10.3778/j.issn.1002-8331.2507-0122
嵌入螺旋感知和多径注意力的轻量化红外目标检测
Lightweight Infrared Object Detection with Spiral Perception and Multi-Path Attention
摘要
Abstract
Aiming at the problems of large difference in target scale,low contrast,less texture information,and large amount of detection model parameters in infrared aerial images,a lightweight infrared target detection model embedded with spiral perception and multipath attention is proposed.Firstly,the spiral convolution is embedded and the global per-ception ability of the model is enhanced by progressive expansion of the receptive field of the convolution kernel.Secondly,in view of the large scale difference between intra-class and inter-class targets,a spatial channel position feature fusion module is designed to strengthen the feature representation ability of infrared multi-scale targets.Then,aiming at the prob-lems of low contrast and less texture information of infrared images,a multi-path collaborative feature enhancement atten-tion module is constructed,and multiple branches are used to capture the features of the target region to improve the focus ability of the model.Finally,the adaptive amplitude pruning method of the application layer is used to dynamically cut the number of model channels,reduce the number of parameters,and enhance the deployment adaptability of the model.The experimental results show that compared with the benchmark method,the detection accuracy of mAP50 and mAP50-95 is improved to 83.8%and 53.3%respectively,the model's GFLOPs is decreased to 7.8,and the inference time is 2.1 ms.By deploying the infrared target detection model on the Jetson Nano mobile development platform,the effectiveness of the practical engineering application is verified,which provides strong support for the infrared monitoring system of urban UAV.关键词
红外航拍检测/螺旋卷积(SPConv)/多径协同特征增强注意力/层自适应幅度剪枝/模型部署Key words
infrared aerial detection/spiral convolution(SPConv)/multipath synergistic feature enhance attention/layer adaptive magnitude pruning/model deployment分类
信息技术与安全科学引用本文复制引用
张麒,何自芬,张印辉,陈光晨,薛金生,王启刚..嵌入螺旋感知和多径注意力的轻量化红外目标检测[J].计算机工程与应用,2026,62(17):283-294,12.基金项目
国家自然科学基金(62561032,62171206,62061022) （62561032,62171206,62061022）
装备智能运用教育部重点实验室开放基金(AAIE-2023-0203) （AAIE-2023-0203）
昆明理工大学分析测试基金(2024P20231103003). （2024P20231103003）
