融合转置卷积的YOLOv3吸烟检测算法

YOLOv3 smoking detection algorithm based on transposed convolution fusion

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点左右。

Abstract:

To prevent safety incidents caused by smoking in public places, an improved smoking detection algorithm is proposed based on the YOLOv3 framework. Firstly, to address the issue of pixel information loss during traditional upsampling operations, a convolutional-transpose convolutional module is designed to replace it. In the feature fusion part, a coordinate attention mechanism is added to make the network better focus on small targets. Improved k-means++ is used to optimize the prior box. Finally, the GIoU is replaced with IoU as the loss function of the algorithm to further improve the detection accuracy. In addition, a multi-scene smoking dataset is constructed, and data augmentation and expansion are performed on the dataset. Experimental results show that the improved algorithm has a 5.58% and 3.34% increase in AP@0.5 and AP@0.5:0.95 compared to the original algorithm, respectively, and the FPS is reduced by about 3 points.