基于改进YOLOv7-tiny的轻量化列车轮对踏面缺陷检测方法

Lightweight train wheelset tread defect detection method based on improved YOLOv7-tiny

摘要:

针对目前列车轮对踏面缺陷检测精度低、检测速率慢和检测类别单一等问题,提出一种改进的轻量化YOLOv7-tiny踏面损伤检测方法。该方法采用轻量级的MobileNetV3网络替代YOLOv7-tiny的主干网络,降低模型的参数量与计算量；将BiFormer注意力机制嵌入主干网络中,可强化学习目标区域的特征,提升模型的检测精度；利用集中式特征金字塔(centralized feature pyramid,CFP)以加强特征的层内调节能力,捕获全局长距离依赖关系与踏面缺陷的局部关键信息；采用WIoU(wise intersection over union)损失函数提升边框回归损失的收敛速度,增强模型的鲁棒性；通过在YOLOv7-tiny的检测头部引入GS解耦头(GSconv decoupled head,GSDH),将分类和回归任务解耦出单独的特征通道,有效地提升模型的并行计算速率和检测精度。实验结果表明:改进的YOLOv7-tiny网络参数量与计算量分别减少11.7%和21.2%,检测准确率提高7.9%,召回率提高10.5%,平均精度均值提高10.1%,检测速度提高7.4帧/s,实现轻量化的同时具有更好的检测性能。改进的方法在轮对踏面损伤检测中具有较广泛的应用前景。

Abstract:

Aiming at the problems of low detection accuracy,slow detection rate and single detection category of wheelset tread defects,an improved lightweight YOLOv7-tiny tread damage detection method is proposed.In this method,the lightweight MobileNetV3 network is used to replace the backbone network of YOLOv7-tiny,and the parameter number and calculation amount of the model are reduced.Embedding BiFormer attention mechanism into the backbone network can strengthen the features of the learning target region and improve the detection accuracy of the model.The centralized feature pyramid (CFP) is used to enhance the feature′s in-layer adjustment ability,capture the global long distance dependence and local critical information of tread defects.Wise intersection over union (WIoU) loss function is employed to accelerate the convergence rate of border regression loss and enhance the robustness of the model.GSconv decoupled head (GSDH) is introduced into the YOLOv7-tiny detection header to decouple separated feature channels from classification and regression tasks,effectively improving the parallel computation rate and detection accuracy of the model.The experimental results show that the improved YOLOv7-tiny network parameter number and computation amount are reduced by 11.7% and 21.2% respectively,the detection precision is increased by 7.9%,the recall rate is increased by 10.5%,the mean average precision is increased by 10.1%,and the frame per second is increased by 7.4 frames/s,which realizes lightweight and has better detection performance.The improved method has a wide application prospect in wheelset tread damage detection.