摘要
针对密集人群下口罩佩戴检测实时性差、难以部署到移动端的问题,提出基于YOLOv7-tiny改进的口罩佩戴检测算法YOLOv7-DSC.该算法结合深度可分离卷积改进的SE注意力机制设计了一种轻量化特征提取模块,并结合BiFPN设计了一种加权特征融合模块.经实验验证,YOLOv7-DSC算法在口罩数据集上mAP为96.9％,与YOLOv7-tiny算法相比仅降低了 0.5％;相比于YOLOv3-ti-ny、YOLOv4-tiny、YOLOv5s、MobileNetV3、ShuffleNetV2、GhostNet 和 Swin-Transformer 算法在 mAP 上分别高出 13.4％、11.2％、4.5％、5.7％、5.8％、4.2％和 5.1％;在检测精度与 YOLOv7-tiny 算法相当的情况下,参数量和计算量分别减少了 60％和55％,仅为2.4 M和6.0 G,极大地降低了硬件成本.
Abstract
This paper proposes an improved mask wearing detection algorithm YOLOv7 DSC based on YOLOv7 tiny to address the problems of poor real-time performance and difficulty in deploying masks to mobile devices in dense crowds.This algorithm combines the SE attention mechanism improved by deep separable convolution to design a lightweight feature extraction module,and combines BiFPN to design a weighted feature fusion module.Through ex-perimental verification,the mAP of YOLOv7-DSC algorithm on the mask data set is 96.9％,which is only 0.5％lower than that of YOLOv7-tiny algorithm.Compared with the YOLOv3-tiny,YOLOv4-tiny,YOLOv5s,Mobile-NetV3,ShuffleNetV2,GhostNet and Swin-Transformer algorithms,the mAP is 13.4％,11.2％,4.5％,5.7％,5.8％,4.2％,and 5.1％higher,respectively;When the detection accuracy is comparable to that of YOLOv7-tiny al-gorithm,the number of parameters and computation are reduced by 60％and 55％,respectively,to only 2.4M and 6.OG,which greatly reduces the hardware costs.
基金项目
国家自然科学基金(61170060)
安徽省重点教学研究项目(2020jyxm0458)
NETL
NSTL
维普
万方数据