A Novel Deep Learning Framework for Recognizing High-Speed Train Numbers

With an upsurge in high-speed railway construction and operation, the demand and necessity for recognizing high-speed train numbers are increasing now. Unlike car license plate numbers, there is no fixed position, color and font for the high-speed train numbers. And during the process of numbers collection, the images about train numbers being photted also are different in the form. However, it is often difficult to obtain high accuracy for high-speed train number recognition using traditional image processing methods. Therefore, this paper proposes a novel deep learning framework, which combines the advantages of three networks, i.e., VGGNet, Faster R-CNN and DenseNet to recognize the high-speed train numbers. The recognition process is divided into two stages: firstly, train numbers are positioned, and secondly train numbers are classified. Among them, the train numbers are positioned by using the updated VGG network for feature extraction, and then region proposals are generated through the Region Proposal Networks in Faster R-CNN. For the classification of high-speed train numbers, the improved DenseNet model is utilized to identify and classify the train numbers images that have been marked with the train numbers area. The experimental results show that the accuracy of this method reaches 98.36%, which is significantly better than other benchmark object detection methods.