基于改进YOLOv7-tiny的安全帽佩戴检测算法

Safety helmet wearing detection algorithm based on improved YOLOv7-tiny

针对目前的安全帽佩戴检测方法在拥挤场景容易漏检和误检小目标和遮挡目标的问题,提出了一种基于YOLOv7-tiny安全帽佩戴检测方法。在YOLOv7-tiny主干网络中加入了协调注意力机制以提高网络对物体位置的感知能力;在特征融合部分设计新的加强特征提取网络结构来实现更高效的特征融合,并使用深度可分离卷积和损失函数SIoU来减少模型的收敛速度和提高检测精度;在预测头上新增一个特征层和检测尺度,进一步加强对小目标的检测能力。将改进的模型命名为DCS-YOLO,通过实验验证了DCS-YOLO模型的有效性,与原始模型相比,DCS-YOLO的模型平均精度达到93.43%,提高了4.67%,同时小目标和遮挡目标的漏检和误检得到了改善,具有良好的检测精度和检测速度,也更容易部署在计算资源和内存有限的设备上。

Aiming at the problem that current helmet wearing detection methods is prone to miss and mis-detect small and obscured targets in crowded scenes,a YOLOv7-tiny helmet wearing detection method is proposed.A coordinated attention mechanism is added to the YoLov7-tiny backbone network to improve the ability of the network to Perceive the location of objects.A new enhanced feature extraction network structure is designed in the feature fusion part to achieve more efficient feature fusion,and the depth-separable convolution and loss function SIoU are used to reduce the convergence speed of the model and improve the detection accuracy;a new feature layer and detection scale on the prediction head to further enhance the detection capability for small targets.The improved model is named DCS-YOLO,and the effectiveness of the DCS-YOLO model is experimentally verified.Compared with the original model,the average model accuracy of DCS-YOLO reaches 93.43%,an improvement of 4.67%,while the missed and false detection of small and occluded targets is improved,with good detection accuracy and detection speed,and is also easier to deploy in devices with limited computational resources and memory-limited devices.