Abstract: In response to the low generalization of traditional railway locomotive number localization and detection models and their inability to adapt to various detection application scenarios, this paper proposed a method for locating and detecting railway locomotive number in unrestricted scenarios based on YOLO (You Only Look Once) v4-tiny model. The paper used cavity convolution instead of standard convolution to increase the receptive field of locomotive number feature extraction, improve the detection accuracy of the traditional YOLOv4 tiny model, established the Railway Locomotive Number Data set (RLND) for model training, and verified the detection effect of the model. The validation results show that the positioning and detection accuracy of this method for railway locomotive numbers is 99.44%, with a detection speed of 50 frames/s. It can meet the needs of locomotive number positioning and detection in unrestricted scenarios.