基于YOLOv4算法的煤矿井下粉尘检测方法

Dust Detection Method for Underground Coal Mines Based on YOLOv4 Algoritm

针对煤矿井下粉尘检测实时性差,容易出现漏检误检的问题,提出一种基于图像的优化YOLOv4煤矿井下粉尘检测算法。该算法将YOLOv4算法与空间变换网络相结合,通过主干特征提取网络得到3个不同尺度的特征图;再送入空间变换网络进行仿射变换,以提高网络的空间变换能力;最后多尺度特征融合网络采用PANet进行特征堆叠,将得到的3个尺度的有效特征层送入特征预测网络进行预测。模型训练阶段对图像进行平移、翻转等预处理,扩充训练数据集,防止检测模型出现过拟合现象。结果表明,优化后的模型对于粉尘图像的检测效果更好,有效降低了目标检测过程中的漏检率。

A YOLOv4-based dust detection algorithm is proposed to address the problems of real-time dust detection in underground coal mines and miss detect. The algorithm combines YOLOv4 with spatial transformation network. Firstly, the feature maps at three different scales are obtained by the backbone feature extraction network;then they are sent to the spatial transformation network for affine transformation to improve the spatial transformation capability of the network;finally, the multi-scale feature fusion network uses PANet for feature stacking, and the obtained effective feature layers at three scales are sent to the feature prediction network for prediction. The improved detection method has translation invariance, scaling invariance and rotation invariance. In the model training stage, image translation and flip are preprocessed to expand the training data set and prevent the detection model from over-fitting. Experiments show that the improved model has better detection effect for dust images compared with the original network, and effectively reduces the leakage rate in the target detection process.