YOLOv5口罩检测算法的轻量化改进  

Lightweight Improvement of YOLOv5 Mask Wearing Detection Algorithm

作　　者：程泽华 韩俊英[1] CHENG Zehua;HAN Junying(School of Information Science and Technology,Gansu Agricultural University,Lanzhou 730070,China)

机构地区：[1]甘肃农业大学信息科学技术学院,甘肃兰州730070

出　　处：《软件导刊》2023年第11期174-179,共6页Software Guide

摘　　要：当前YOLOv5算法在口罩识别检测任务中,当光线昏暗时会出现个别目标无法正确识别,同时也存在网络参数量大、在移动设备上运行速度慢的问题。对比通过将空间可分离卷积和深度可分离卷积进行自适应特征融合,设计一种融合注意力的SDC模块,其减少了网络模型的参数量,增强了网络的特征提取能力,提高了模型检测精度。使用随机遮挡和自适应对比度的数据增强方法,模拟了实际场景中存在目标遮挡或光线昏暗等现象。选用SIoU Loss作为边框回归损失函数,提高了网络训练速度和推理准确性。实验结果表明,相较于原始的YOLOv5算法,其平均精度均值(mAP)提高了1.3个百分点,检测帧率(FPS)提升约140%,能够更好地满足口罩检测任务的准确性和实时性。In the current mask recognition and detection task of YOLOv5 algorithm,when the light is dim,individual targets can not be correctly identified,and there are also problems such as large network parameters and slow running on mobile devices.Through adaptive feature fusion(AFF)of space separable convolution and depth separable convolution,a SDC(spatial deep convolution)module integrating attention is designed,which reduces the parameters of the network model,enhances the network feature extraction ability,and improves the detection accuracy of the model.The data enhancement methods of random occlusion and adaptive contrast are used to simulate the phenomena of occlusion or dim light in real scenes.SIoU Loss is selected as the frame regression loss function to improve the training speed and reasoning accuracy of the network.The experimental results show that compared with the original YOLOv5 algorithm,the mean precision(mAP)is improved by 1.3 percentage points,and the detection frame rate(FPS)is improved by about 140%,which better meets the accuracy and real-time performance of the mask detection task.

关 键 词：口罩检测 检测帧率 SIoU Loss YOLOv5 光线昏暗 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]