面向航拍图像的轻量化目标检测算法  

A Lightweight Target Detection Algorithm for Aerial Images

作　　者：何戚天 李为相 程明 孙圆 陈闯 HE Qitian;LI Weixiang;CHENG Ming;SUN Yuan;CHEN Chuang(College of Electrical Engineering and Control Science Nanjing Tech University,Nanjing 211000 China)

机构地区：[1]南京工业大学电气工程与控制科学学院,南京211000

出　　处：《电光与控制》2025年第3期56-61,81,共7页Electronics Optics & Control

基　　金：国家自然科学基金(62303217);江苏省高等学校基础科学(自然科学)研究项目(23KJB510006)。

摘　　要：针对无人机航拍图像背景复杂、目标物较小且模型部署困难的问题,提出了面向航拍图像的轻量化目标检测算法。在YOLOv5m的主干网络中引入FasterNet轻量化模块替换C3模块,压缩模型参数量,提升模型的推理速度;在特征融合网络,采用改进的CBAM_L机制,专注于捕捉航拍图像中的小目标信息的同时提高了模型的目标识别精度;将检测网络中检测头替换成解耦头,解决航拍图像输出变量时分类和回归的冲突问题,并把网络中损失函数替换成EIoU,有效提升了模型回归精度。在公开数据集VisDrone上验证结果表明,改进后模型的平均精度均值(mAP@0.5)提高了0.014,参数量缩减到原模型的34.3%,计算量缩减到原模型的32.4%,检测速率达到77帧/s,表明该算法在检测准确性和速度上都取得出色的性能。In order to solve the problems of complex background small target objects and difficult model deployment in UAV aerial images a lightweight target detection algorithm for aerial images is proposed.The lightweight FasterNet module is introduced in the backbone network of YOLOv5m to replace C3 module and the model parameters is compressed to improve reasoning speed of the model.In the feature fusion network the improved CBAM_L mechanism is used to focus on capturing small target information in aerial images while improving the target recognition accuracy of the model.The detection head in the detection network is replaced by a decoupled head which solves the conflict between classification and regression when outputting variables in aerial images;and the loss function in the network is replaced by EIoU which effectively improves the model regression accuracy.The verification results on the public dataset VisDrone show that the average accuracy mAP@0.5 of the improved model is increased by 0.014 the parameter quantity and the computation cost is respectively reduced to 34.3%and 32.4%of the original model and the detection speed reaches 77 frames per second.The results show that the proposed algorithm exhibits good performance in both detection accuracy and speed.

关 键 词：YOLOv5m FasterNet 轻量化 注意力机制 解耦头 EIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]