基于改进YOLOv5的轻量级落头目标检测方法  

Lightweight Fallout Target Detection Method Based on Improved YOLOv5

作　　者：张浩 牛芳芳[2] 桑瑶烁[2] ZHANG Hao;NIU Fangfang;SANG Yaoshuo(School of Electronic and Information Engineering,Anhui Jianzhu University,Hefei 230009;Hefei Institutes of Physical Sciences,Chinese Academy of Sciences,Hefei 230031)

机构地区：[1]安徽建筑大学电子与信息工程学院,合肥230009 [2]中国科学院合肥物质科学研究院,合肥230031

出　　处：《计算机与数字工程》2025年第2期444-450,共7页Computer & Digital Engineering

摘　　要：针对燃烧锥落头的准确快速检测需求,提出一种基于改进YOLOv5的轻量级落头检测方法。首先,使用Ghost模块替换Backbone和Neck中的Conv模块和C3模块,提升落头的检测速度;其次,在Neck部分中添加一条加权双向特征金字塔网络BiFPN,通过加权来更好地平衡不同尺度的特征信息改善检测精度;最后,使用α-IoU损失函数替换GIoU损失函数优化边界框回归损失函数,获得更准确的边界框定位精度。实验结果表明:改进的YOLOv5算法的落头识别平均准确率达到了96.1%,mAP0.5和mAP0.5:0.95分别达到了98.3%和73.0%,同时,模型权重大小仅为8.7 M,能够满足燃烧锥落头准确快速的检测要求。To address the equipment of accurate and fast detection of fallout,a lightweight fallout detection method based on improved YOLOv5 is proposed.Firstly,the Conv module and C3 module in Backbone and Neck are replaced by Ghost module to im⁃prove the detection speed of fallout.Secondly,a weighted bidirectional feature pyramid network BiFPN is added to the Neck part to improve the detection accuracy by better balancing the feature information of different scales through weighting.Finally,the GIoU loss function is replaced byα-IoU loss function to optimize the bounding box regression loss function to obtain more accurate bound⁃ing box localization accuracy.The experimental results show that the average accuracy of the improved YOLOv5 algorithm for fallout recognition reaches 96.1%,and the mAP0.5 and mAP0.5:0.95 reach 98.3%and 73.0%,respectively.Meanwhile,the model weight size is only 8.7 M,which can meet the requirements of accurate and fast detection of fallout.

关 键 词：YOLOv5 落头检测 轻量化 BiFPN 损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]