关键词: 铁路客车, 缺陷图像, 目标检测, Faster R-CNN, K-means++

Abstract: In order to ensure operation safety of railway passenger cars, an image defect detection algorithm for key parts of them based on Faster R-CNN target detection is proposed together with two points for improvement considering the algorithm's problems in detecting small-scale bolt targets. Firstly, original VGG16 network was replaced by advantages of deep residual network and Inception network, and upsampling layer was added to solve serious loss of image feature information through convolutional network. Secondly, size and proportion of anchor points in Region Proposal Network (RPN) were optimized through K-means++ clustering algorithm to improve accuracy of generated suggested regions and address inaccurate positioning of defective targets. Finally, comparative experiments on bolt defect dataset created in this paper were conducted. The results show that detection accuracy of improved algorithm reaches 87.4%, which is 8.9% higher than original one. Moreover, its missed detection rate and false detection rate are reduced by 9.9% and 11% respectively for multi-target defects and confused targets.

Key words: railway passenger cars, defect image, target detection, Faster region-convolutional neural network(Faster R-CNN), K-means++