Improving Yolov6 Mine Object Detection

In order to solve the problem of traditional convolutional networks detecting multiple targets missing underground and inaccurate detection of small obstacles, we propose an improved YOLOv6 algorithm aimed at improving the safety of mine employees. Firstly, we combine the multi head self attention mechanism with the maximum pooling layer to replace a portion of the convolutional layer in the model backbone framework, thereby improving the learning ability of the convolutional network for the global features of the target object. The network structure has been improved in the bottleneck layer, retaining the original detection accuracy while improving the target detection ability. In order to expand the receiving field range of the convolution layer, we use the funnel Activation function FRelu as the Activation function of the backbone network. In addition, we introduce depth separable convolution to improve the Rate of convergence of the model. After experimental verification, the improved detection algorithm has achieved significant results in mine detection. This algorithm determines the safety of employees in complex mine environments through portrait recognition in practical engineering applications.