现代电子技术2024，Vol.47Issue(1)：44-50,7.DOI:10.16652/j.issn.1004-373x.2024.01.008
基于轻量级网络的小目标检测算法
Small object detection algorithm based on lightweight network
摘要
Abstract
In view of the low accuracy of YOLOv5 algorithm in detecting small objects,a lightweight network KOS-YOLOv5 which aims to improve the accuracy of small object detection is proposed.The K-means++ clustering technology is used to select a set of appropriate anchor box sizes as priors for the model,achieving more accurate anchor box sizes for small objects and enabling the model to adapt to objects with different sizes.A simplified positive-negative sample assignment strategy,which refers to SimOTA,is utilized for dynamic sample matching,so as to optimize the loss function more effectively.The spatial context pyramid(SCP)module is integrated into the algorithm's detection layer,which encourages the backbone network to pay more attention to the feature information of small objects,so as to enhance the capability of extracting object features and improve the accuracy of object detection.The results show that the improved KOS-YOLOv5 algorithm improves the detection precision by 4%,the recall rate by 2.4%,the mean average precision by 3.1%and the loss function value by 5%in comparison with those of the traditional YOLOv5 model.The final detection accuracy of the improved algorithm is 95.38%.关键词
小目标检测/轻量级网络/特征提取/优化损失函数/YOLOv5/K-means++Key words
small object detection/lightweight network/feature extraction/optimized loss function/YOLOv5/K-means++分类
电子信息工程引用本文复制引用
关玉明,王肖霞,杨风暴,吉琳娜,丁春山..基于轻量级网络的小目标检测算法[J].现代电子技术,2024,47(1):44-50,7.基金项目
国家自然科学基金项目(61972363) （61972363）
中央引导地方科技发展资金项目(YDZJSX2021C008) （YDZJSX2021C008）
中北大学研究生科技立项项目(20221832) （20221832）
