基于改进YOLOv5的煤矿井下目标检测研究

针对煤矿井下环境多利用红外相机感知周边环境温度成像，但形成的图像存在纹理信息少、噪声多、图像模糊等问题，该文提出一种可用于煤矿井下实时检测的多尺度卷积神经网络(Ucm-YOLOv5)。该网络是在YOLOv5的基础上进行改进，首先使用PP-LCNet作为主干网络，用于加强CPU端的推理速度；其次取消Focus模块，使用shuffle_block模块替代C3模块，在去除冗余操作的同时减少了计算量；最后优化Anchor同时引入H-swish作为激活函数。实验结果表明，Ucm-YOLOv5比YOLOv5的模型参数量减少了41%，模型缩小了86%，该算法在煤矿井下具有更高的检测精度，同时在CPU端的检测速度达到实时检测标准，满足煤矿井下目标检测的工作要求。

Research on Target Detection in Underground Coal Mines Based on Improved YOLOv5

In view of the underground coal mine environment, which uses mostly infrared cameras to sense the surrounding environment’s temperature, the images formed have the problems of less texture information, more noise, and blurred images. The detection of Underground targets in coal mines using YOLOv5(Ucm-YOLOv5), a neural network for real-time detection of coal mines, is suggested in this document. This network is an improvement on YOLOv5. Firstly, PP-LCNet is used as the backbone network for enhancing the inference speed on the CPU side. Secondly, the Focus module is eliminated, and the shuffle_block module is used to replace the C3 module in YOLOv5, which reduces the computation while removing redundant operations. Finally, the Anchor is optimized while introducing H-swish as the activation function. The experimental results show that Ucm-YOLOv5 has 41% fewer model parameters and an 86% smaller model than YOLOv5. The algorithm has higher detection accuracy in underground coal mines, while the detection speed at the CPU side reaches the real-time detection standard, which meets the working requirements for target detection in underground coal mines.