Abstract: Concerning the problems of large scale change and small sample dataset in rail tread block defects, a rail tread block defects detection method based on improved Faster Region-based Convolutional Neural Network (Faster R-CNN) was proposed. Firstly, based on the basic network structure of ResNet-101, a multi-scale Feature Pyramid Network (FPN) was constructed to achieve the fusion of deep and shallow feature information in order to improve the detection accuracy of small-scale defects. Secondly, the Generalized Intersection over Union (GIoU) loss was used to solve the problem of insensitivity to the position of the predicted border caused by regression loss SmoothL1 in Faster R-CNN. Finally, a method of Region Proposal Network by Guided Anchoring (GA-RPN) was proposed to solve the problem of the imbalance of positive and negative samples in the training of the detection network due to the large redundancy of anchor points generated by Region Proposal Network (RPN). During the training process, the RSSDs dataset was expanded based on image preprocessing methods such as flipping, cropping and adding noise to solve the problem of insufficient training samples of rail tread block defects. Experimental results show that the mean Average Precision (mAP) of the rail tread block defects detection based on the proposed improved method can reach 82.466%, which is increased by 13.201 percentage points compared with Faster R-CNN, so that the rail tread block defects can be detected accurately by the proposed method.

Key words: rail tread, block defects detection, Faster Region-based Convolutional Neural Network (Faster R-CNN), Feature Pyramid Network (FPN), Generalized Intersection over Union (GIoU), Region Proposal Network (RPN)

摘要： 针对钢轨踏面块状伤损存在的尺度变化大、样本数据集小的问题，提出了基于改进Faster R-CNN的钢轨踏面块状伤损检测方法。首先，基于ResNet-101基础网络结构来构建多尺度特征金字塔（FPN），以实现深、浅层特征信息的融合，从而提高了小尺度伤损的检测精度；然后，采用广义交并比（GIoU）损失解决了Faster R-CNN中回归损失SmoothL1对预测边框位置不敏感问题；最后，提出引导锚定的区域提名网络（GA-RPN）方法，从而解决了区域生成网络（RPN）生成的锚点大量冗余而导致的检测网络训练中正负样本失衡问题。训练过程中，基于翻转、裁剪、噪声扰动等图像预处理方法对RSSDs数据集进行扩充，解决了钢轨踏面块状伤损训练样本不充足问题。实验结果表明，所提改进方法对钢轨踏面块状伤损检测的平均精度均值（mAP）可达到82.466%，相较于Faster R-CNN提高了13.201个百分点，能够更加准确地检测钢轨踏面块状伤损。

关键词: 钢轨踏面, 块状伤损检测, Faster区域卷积神经网络, 特征金字塔, 广义交并比, 区域建议网络