Abstract: Aiming at the problems of low accuracy and high missed detection rate of personnel safety helmet detection algorithm caused by complex environment under mine, this paper proposes an improved mine safety helmet detection algorithm based on YOLOv5 s. Due to the computer system, the global context information of the image is easily lost when the convolutional neural network extracts the features, resulting in poor detection effect of the downhole small target safety helmet. To this end, this paper uses the attention mechanism CBAM and YOLOv5s to fuse, enhance the feature map of the target area, and weaken the background information, so as to help the algorithm better locate the small target helmet. At the same time, a P2 small target detection layer is added on the basis of the original three output layers of YOLOv5 s, which increases the multi-scale receptive field of the model and can capture global and local context information at the same time, which improves the detection ability of the algorithm for small targets in complex scenes. In addition, this paper uses EIoU loss to replace the original CIoU loss function to solve the problem of fuzzy width-to-height ratio of the prediction frame, ensure the accuracy of the regression frame, and accelerate the convergence speed of the network. By replacing the ordinary convolutional Conv in the YOLOv5s backbone network with ShuffleNetV2, the number of model parameters is greatly reduced and the recognition speed of the model is improved. Finally, the proposed algorithm is compared with YOLOv5 s, SSD, Faster-RCNN and YOLOv7 algorithms. The experimental results show that the improved algorithm is applied to the safety helmet detection of mine personnel. Compared with the original YOLOv5 s, the accuracy of the proposed algorithm is increased by 2.9%, the recall rate is increased by 2.42%, and the parameter quantity is reduced by 7.6%. Finally, the average accuracy mAP@.5 of the algorithm in the mine safety helmet detection reaches 87.5%.