Wagon Number Recognition Based on the YOLOv3 Detector

The automatic recognition of wagon number plays an important role in the railway transportation system, but high-precision recognition is still challenging due to many disturbing factors, such as uneven illumination, complex background, image fouling and various type of wagon number. To reduce the complexity of the problem, we design a two-stage approach for wagon number recognition based on the state-of-the-art object detector YOLOv3. Firstly, the region of wagon number is detected by one detector and cropped according to the predicted bounding box, then the digits that make up the wagon number are detected and recognized in the cropped image by another detector. During the whole process, there is no need for image preprocessing and character segmentation, which reduce the accumulation of intermediate error and improve the speed of recognition. Experimental results show that the approach achieves a recognition rate of 96.08% and speed of 70ms per image on 1072 test images.