基于半监督深度学习的无砟轨道扣件缺陷图像识别方法

Image Recognition Method for the Fastener Defect of Ballastless Track Based on Semi-Supervised Deep Learning

提出基于置信图的扣件子图快速定位算法,即基于概率图模型构建扣件子图邻域纹理图和初始引导图与被定位图之间的置信图,并通过计算置信图的最大极值点实现对扣件子图中心点的快速定位。在此基础上,针对无砟轨道扣件缺陷样本相对稀缺的问题,提出基于半监督深度学习的扣件缺陷图像识别方法,即首先采用稀疏自编码(SAE)网络在无标签的数据集进行迭代学习获得扣件子图稀疏表征,然后将训练好的SAE网络连接softmax层组成分类网络,最后在有人工类别标注的小数据集进行二次训练及参数微调获得最终的识别模型。通过在装配WJ-7型扣件的CTRS-Ⅰ和WJ-8型扣件的CTRS-Ⅱ型无砟轨道图像进行应用测试和方法验证。结果表明:该方法可快速精确定位扣件并识别扣件缺失、弹条折断、弹条移位3类缺陷,有效检出率达95%以上。

A fast location algorithm for fastener sub-image based on confidence map was proposed.Namely,the neighborhood texture map of fastener sub-image as well as the confidence map between the initial boot map and location map was generated based on probabilistic graphical model.The maximum extreme point of the confidence map was calculated to quickly position the center point of fastener sub-image.On this basis,a recognition method for fastener defect image was proposed based on semi-supervised deep learning to overcome the difficulty in relatively small number of fastener defect samples of ballastless track.The sparse representation of fastener sub-image was first obtained by the iterative learning of the Sparse Auto-Encode（SAE）network in the unlabeled datasets.Then the trained SAE network was connected to the softmax layer to build the classification network.Finally,the final recognition model was obtained by the second training and parameter tuning in the small datasets with manual category labels.Application test and method validation were carried out on the images of the CTRS-I ballastless track with WJ-7 fasteners and CTRS-II ballastless track with WJ-8 fasteners.Results show that the method can quickly and accurately locate fasteners and recognize three kinds of defects,such as the missing of fastener,the break and shift of bar-spring clips.The effective detection rate is above 95%.