YOLOv3 Sleeper Crack Detection Algorithm with Attention Mechanism
ZHU Jiuniu
LI Liming
ZHENG Shubin
PENG Lele
CHAI Xiaodong
Abstract：In response to the poor performance of current methods such as manual inspection and physical equipment-assisted nondestructive testing in rail sleeper crack detection,this paper proposes an improved YOLOv3 model.This paper proposes an im-proved backbone feature extraction network fused with coordinate attention mechanism,which enables it to capture the long-range dependence of spatial information and thus better locate the crack location.The RFN module proposed in this paper is able to en-hance the deep semantic information and fuse the global information of the overall network before performing feature fusion.The co-ordinate regression loss using DIoU instead of YOLOv3 makes the network converge earlier.Experiments show that the improved YO-LOv3 has better detection performance,and the improved target detection algorithm proposed in this paper is tested on the same rail crack dataset for both,in which the average accuracy is improved by about 5.2%on average,the precision is improved by about 2.5%,and the recall is improved by about 7.2%.The ablation experiments designed in this paper show that the use of SPP module as well as RFN module improves the detection accuracy of the model,and the use of DIoU loss function not only accelerates the con-vergence of the network,but also provides higher recall and detection accuracy.Overall,the model is able to perform the task of rail sleeper crack detection with high speed and high accuracy.
Keywords：sleeper crack detectionYOLOv3 improvementRFN modulespatial pyramid poolingDIoU loss function
Publication Date：2025-12-20
Online Publishing Date：2026-03-09(First online date of this platform, not the publication date of the document)
Pages：6( 3332-3336,3406 )