Abstract: In order to solve the problems of low detection accuracy and slow detection speed of small defects on complex rail surface by traditional methods, a rail surface defect detection algorithm based on multi-perception synergy and mixed sampling strategy was proposed. Firstly, an improved lightweight feature extraction backbone CASG-MobileNetV2 was constructed to realize the lightweight model and effectively enhance the ability of the lightweight backbone to extract defect features. Secondly, a pyramid module of foreground perception attention collaborative features is proposed to extract multi-dimensional defect features in complex orbit scenes to enhance the detection effect of small targets. Then, a hybrid sampling strategy was designed in the Transformer part to replace the self-attention learning with dynamic perception, so as to reduce the computational cost of the model and further capture the global and local feature information. Finally, the defect detection output is completed by the feedforward neural network and the Hungarian matching algorithm. Experimental results show that the proposed algorithm is increased by 3.5% points to 71.3% compared with the mean Average MAPsion (mAP@0.5) of the original DETR model, the parameter amount is compressed by 44.5%, and the detection rate is increased to 43.7 frames/s, which is 1.6 times that of the original algorithm. The evaluation index of the proposed method is better than that of the comparison method, and it can quickly and accurately detect the surface defects of the rail.

Key words: railway engineering；railway transportation; defect detection; DETR; lightweight inspection; mixed sampling

摘要： 针对传统方法对复杂钢轨表面较小缺陷检测精度低、检测速度慢等问题，提出一种多元感知协同与混合采样策略的钢轨表面缺陷检测算法。首先，构建改进的轻量级特征提取主干CASG-MobileNetV2，实现模型轻量化的同时，有效增强轻量级主干对缺陷特征的提取能力；其次，提出前景感知注意力协同特征金字塔模块，在复杂轨道场景中进行多维度缺陷特征提取，增强小目标检测效果；再次，在Transformer部分设计混合采样策略，以动态感知代替自注意力学习，从而降低模型计算量，并进一步捕获全局与局部特征信息；最后，通过前馈神经网络与匈牙利匹配算法完成缺陷检测输出。实验结果表明：笔者算法较原始DETR算法平均精度均值（mAP@0.5）提升了3.5%，达71.3%；参数量压缩44.5%；检测速率提升至43.7帧/s，为原始DETR算法的1.6倍。笔者算法的评价指标优于对比方法，能够快速准确地检测出钢轨表面缺陷。

关键词: 铁道工程； 铁路运输；缺陷检测；DETR；轻量级检测；混合采样