Research on segmentation of key parts for train safety based on deep learning

Abstract

Image segmentation is the most fundamental part of computer vision, which is the foundation of all other methods of image processing. The quality of image segmentation technology will affect the subsequent processing considerably. Comparing with traditional image segmentation algorithms, image segmentation algorithm based on deep learning is constantly proposed, with high performance and efficiency. But there is also a lot of room for improvement. For example, key parts such as fastening bolt are usually small in size, polluted and covered, and do not have enough characteristic information, so it is difficult to obtain satisfactory results. These factors affect the accuracy of the test, which is easy to cause serious accidents. As traditional methods sometimes cannot meet the requirement of high-accuracy result, deep learning play a particularly important role in facing those problems. To solve the problem that traditional object recognition methods are not robust enough to extract image features, parts recognition accuracy is low, and segmentation is not possible, we have made some modifications based on Mask R-CNN. In this method, convolutional neural network is used to extract features from part images. Then we use some annotated images from dataset to fine-tuned Mask R-CNN network to guarantee the accuracy. At the same time, data enhancement and k-folding cross-validation are carried out to improve the robustness of the model. Finally, the result of part recognition and segmentation by building the experimental platform proves the significance of the method.