【摘要】 针对矿工安全帽检测算法准确度不高的问题，提出一种基于YOLOv7-tiny网络改进的安全帽检测算法。在YOLOv7-tiny网络的基础上，首先针对安全帽这种小目标检测问题，颈部网络融合了160×160大小特征层的安全帽特征信息，在密集情况下对安全帽的识别效果更好；接着在主干网和颈部网络中间加入SimAM无参注意力机制，有效增强安全帽特征提取能力，丰富模型捕获的上下文信息；最后使用SIoU损失函数替代CIoU损失函数，减小损失自由度，增强网络的鲁棒性。在矿工安全帽数据集上的实验结果表明，改进YOLOv7-tiny算法的平均准确度均值达到98.47%,较YOLOv7-tiny提升3.86%,较YOLOv5-s、YOLOX-tiny目标检测算法分别提升2.90%、7.20%,检测效果优于同类型网络检测效果，在边缘端设备JetsonTX2上检测速度达到28.04帧/s,满足实时性要求。在不同光线场景下的矿工安全帽检测效果显示，改进YOLOv7-tiny算法漏检和误检情况更少，进一步说明了改进算法的有效性。更多还原

【Abstract】 Facing the problem of low accuracy of miners’ safety helmet algorithm, an improved safety helmet detection algorithm based on YOLOv7-tiny network was proposed. Firstly, for the small target detection problem of safety helmet, the neck network integrated the safety helmet feature information of the 160×160 size feature layer on the basis of Yolov7-tiny network, which had a better recognition effect on helmet feature in dense situations. Secondly, SimAM nonparametric attention mechanism was added between the backbone network and the neck network, which effectively enhanced the ability to extract helmet features and enriched the context information captured by the model. Finally, CIoU loss function was replaced by SIoU loss function to reduce the degree of freedom of loss and enhance the robustness of the network. The experimental results on the miners’ helmet dataset show that the average accuracy of the improved YOLOv7-tiny algorithm is 98.47%, 3.86% higher than YOLOv7-tiny. Compared with YOLOV5-s and YOLOX-tiny target detection algorithms, the average accuracy rate increases by 2.90% and 7.20%, respectively. The detection effect is superior to the same type of network detection effect. The detection speed on the edge device JetsonTX2 reaches 28.04 frames per second, meeting the real-time requirements. The detection results of the miner’s helmet in different light scenes show that the missed and false detections of the improved YOLOv7-tiny algorithm are less, which further illustrates the effectiveness of this improved algorithm.更多还原