基于改进YOLOv5的安全帽佩戴检测

针对YOLOv5无法通过权重进行聚焦，产生更具有分辨性的特征，从而降低安全帽检测准确性的问题，使用注意力模块，并分别研究压缩激励层（Squeeze and Excitation Layer, SEL）和高效通道注意力（Efficient Channel Attention, ECA）模块。针对YOLOv5去除冗余框时采用的非极大值抑制（Non Maximum Suppression, NMS）在物体高度重叠时仅保留同类最高置信度预测框的问题，使用Soft-NMS算法保留更多的预测框，并进一步使用加权非极大值抑制（Weighted Non Maximum Suppression, WNMS）融合多次预测框信息提升预测框准确性；针对下采样带来的信息丢失问题，使用Focus模块提升检测效果；综合各个模块得到最优的FESW-YOLO算法。该算法在安全帽数据集上的mAP@0.5、mAP@0.5:0.95相较于YOLOv5分别提高了2.1个百分点、1.2个百分点，提升了安全帽监管准确性。

Helmet-wearing Detection Based on Improved YOLOv5

To the problem that YOLOv5 cannot be focused by weights and cannot produce more distinguishable features, thereby reducing the accuracy of helmet detection, attention module was used. Besides, squeeze and excitation layer and efficient channel attention module were studied. To the problem that the non maximum suppression used by YOLOv5 to remove redundant results will only retain the highest confidence prediction frame of the same class when objects were highly overlapped, the Soft-NMS algorithm was used to keep more prediction boxes. Weighted non maximum suppression was used to fuse multiple prediction boxes information to improve the accuracy of the prediction boxes. For the problem of information loss caused by down-sampling , focus modules was used to improve the detection effect, and the various modules were integrated to obtain the optimal FESW-YOLO algorithm. Compared with YOLOv5, the algorithm improves the mAP@0.5 by 2.1 percentage points and the mAP@0.5:0.95 by1.2 percentage points on the helmet data set respectively, which improves the accuracy of safety helmet supervision.