Aiming at the small target characteristic of rail surface crack and the low precision and slow speed of traditional detection methods,we propose an object detection method based on improved YOLOV4 network for cracks on the surface of rails in this paper.Firstly,in order to obtain the larger effective receptive field area of the feature map and improve the detection accuracy,we use the improved receptive field block (RFB) module to replace the spatial pyramid pooling (SPP) structure;Secondly,we use the deep separable convolution structure to replace the common convolution structure in the network model,so that the network is lightweight and the detection speed is improved;At the same time,we use K-means + + algorithm to reacquire the anchor frame, and then change the linear scale of the anchor frame to solve the problem that the original anchor frames are not suitable for small target detection.The results show that the mean average precision (mAP) of the improved YOLOV4 is 84.8%,which is 3.4% higher than that of the original YOLOV4 algorithm;The detection speed (FPS) is 62.39 frame/s,which increases by 4.07 frame/s.