融合转置卷积的YOLOv3吸烟检测算法
YOLOv3 Smoking Detection Algorithm Fused with Transposed Convolution
龚英杰 1沈希忠1
作者信息
- 1. 上海应用技术大学 电气与电子工程学院,上海 201418
- 折叠
摘要
为预防公共场所因吸烟而引发的安全事故,在YOLOv3框架的基础上提出了改进的吸烟检测算法;首先针对传统上采样操作丢失像素信息等问题,设计出一种卷积-转置卷积模块进行替换;在特征融合部分加入坐标注意力机制,使网络更好关注小目标;使用改进的k-means++优化先验框;最后将GIoU替换IoU作为算法的损失函数,进一步提高检测精度;此外,构建了一个多场景的抽烟数据集,并对数据集进行数据增强与扩充;实验结果表明,改进后的算法较原算法在AP@0.5和AP@0.5∶0.95上分别提高5.58％和3.34％,FPS降低3左右.
Abstract
To prevent safety incidents caused by smoking in public places,an improved smoking detection algorithm based on YOLOv3 framework is proposed.Firstly,to address the missing pixels in traditional upsampling,a convolution module with convolu-tional transpose for the replacement is designed,In the feature fusion part,the coordinate attention mechanism is added to make the network better focus on small targets.The improved k-means++is used to optimize the prior box.Finally,the generalized intersec-tion over union(GIoU)is replaced with the intersection over union(IoU),it is taken as the loss function of the algorithm to further improve the detection accuracy.In addition,the multi-scene smoking dataset is constructed to achieve the data augmentation and ex-pansion on the dataset.Experimental results show that compared to the original algorithm,the improved algorithm increases the AP@0.5 and AP@0.5∶0.95 by 5.58％and 3.34％,respectively,and the frames per second(FPS)decreases by about 3 points.
关键词
深度学习/目标检测/小目标/吸烟/转置卷积/注意力机制Key words
deep learning/target detection/small target/smoking/transposed convolution/attention mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据