A Multi-stage Optimization based Model for Component Defect Detection of Trains

Long runs of railway trains inevitably lend to fault conditions. Automatic detection of defects in train components is very important for train safety and maintenance. For a long time, our detection of images still required to some extent human visual interpretation. ln recent years, convolutional neural networks have achieved great success in object detection, instance segmentation and image classification. In order to reduce the need of manual inspection, in this paper a multi-level optimized instance detection model (MOIDM) is proposed for instance-level defect detection. In MOIDM, we propose a two-stage instance segmentation method for defect detection. First, the feature pyramid network and the depth residual neural network are used to extract the multi-scale features of the image, and then the region proposal neural network is used to generate the rough defect region with the extracted hierarchical features, and the defect classification is carried out. Finally, the obtained features are used for instance-level prediction of coarse defect areas. In addition, we propose an Improved Pyramid Pooling Module (IPPM) and a multi-level optimisation module CPSPMask to improve the performance. IPPM can effectively establish long-range connections between pixels, further increasing the network’s ability to extract feature and spatial context information. CPSPMask can aggregate the original image and the mask for obtaining a more accurate mask. Experiment on a dataset containing two kinds of defect verifies the effectiveness of the algorithm. Our method can achieve 86.8% accuracy in terms of segmentation. In addition, which proves its promising performance comparison with state-of-the-art methods.