Abstract
The contact curve can be extracted from the wheel-rail contact image to better study the wheel-rail contact relationship. To solve this problem, an improved Deeplabv3 + semantic segmentation model is proposed. Firstly, the feature extraction network in the original Deeplabv3 + model was optimized to lightweight MobileNetv2 to reduce the number of model parameters. Secondly, the original parallel atrous spatial pyramid pooling was optimized into a closely connected structure to improve the network extraction capability. Finally, the dual attention mechanism was added after the backbone network was extracted to improve the accuracy of the model. Experiments on wheel-rail contact image dataset showed that the improved model has good real-time performance, and the mean pixel accuracy is 88.72%, the mean intersection over union is 85.84%, and the F1-score is 91.48%, which are 2.29%, 2.63% and 2.15% higher than the original model, respectively. The method can be used as an effective method for wheel-rail contact image segmentation.