首页|基于坐标注意力和软化非极大值抑制的密集安全帽检测

基于坐标注意力和软化非极大值抑制的密集安全帽检测

扫码查看
为解决现有的安全帽检测算法对密集小目标的检测精度低的问题,提出一种基于坐标注意力和软化非极大值抑制的安全帽检测算法.引入坐标注意力机制,聚焦训练安全帽相关目标特征以提高准确率.采用软化非极大值抑制算法对候选框的置信度进行优化,提升模型对密集小目标的检测精度.通过WIoU优化边界框损失函数,使得模型聚焦于困难样例而减少简单示例对损失值的贡献,提升模型的泛化性能.实验结果表明:与基准模型YOLOv5s相比,所提算法的mAP@0.5达到88.4%,提升了3.0%;mAP@0.5:0.95达到65.6%,提升了6.8%;在召回率和准确率上分别提升了2.4%和0.5%.所提算法为密集小目标的检测提供了一定参考.
Dense safety helmet detection based on coordinate attention and soft NMS
In order to solve the problem of low detection accuracy of dense small targets in existing helmet detection algorithms,a safety helmet detection algorithm based on coordinate attention and soft non-maximum suppression(NMS)is proposed.A coordinate attention mechanism is introduced to improve accuracy by focusing on target features associated with the training safety helmet.The softened non-maximum suppression algorithm is used to optimize the confidence of the candidate box,so as to increase the detection precision of the model for dense small targets.The WIoU is used to optimize the bounding box loss function,which can make the model focus on difficult examples and reduce the contribution of simple examples to the loss value,improving the generalization performance of the model.The experimental results indicate that,in comparison with the standard YOLOv5s,the mAP@0.5 of the proposed algorithm is 88.4%,which is increased by 3.0%,the mAP@0.5:0.95 is 65.6%,which is increased by 6.8%,and the recall rate and accuracy are increased by 2.4%and 0.5%,respectively,which can provide a certain reference for the detection of dense small targets.

safety helmet detectioncoordinate attention mechanismSoft non-maximum suppressionYOLOv5sWIoUbounding box loss function

尹向雷、苏妮、解永芳、屈少鹏

展开 >

陕西理工大学 电气工程学院,陕西 汉中 723000

安全帽检测 坐标注意力机制 软化非极大值抑制 YOLOv5s WIoU 边界框损失函数

2025

现代电子技术
陕西电子杂志社

现代电子技术

北大核心
影响因子:0.417
ISSN:1004-373X
年,卷(期):2025.48(2)