Recognizing Train Brake Shoes Based on the Improved OCRNet
Chengli Zhong · 2023
The brake shoe is a key component used for braking in train operation; once the brake shoe is severely worn or falls off, it may cause a major traffic accident. Therefore, the accurate inspection of locomotive brake shoes is of great importance. An improved semantic segmentation algorithm based on an object contextual representations network (OCRNet) is proposed to recognize brake shoes. Firstly, for the problem of brake shoe contour detail segmentation, a coordinate attention mechanism module (CA Block) is introduced into the backbone network of OCRNet to capture the internal correlation of position information and channel information, which can overcome not only the information loss caused by the downsampling process but also improve the semantic segmentation of brake shoe accuracy. In addition, semantic segmentation only focuses on pixel-level classification but lacks the optimization of object-level classification. So, we use a mixed loss function composed of cross-entropy loss and Lovász-softmax loss to supervise the training process at the pixel and object levels, respectively. Finally, the classification network is introduced to improve the detection accuracy and reduce the false positive rate. The proposed method is verified on the test set, and the experimental results show that the improved improved OCRNet can achieve higher accuracy with a slight increase in parameters and calculation amount.