中文 | English

J. Hefei Univ. Tech. (Nat. Sci.)

合肥工业大学学报(自然科学版)

JOURNAL OF HEFEI UNIVERSITY OF TECHNOLOGY (NATURAL SCIENCE)

Competent Authorities: Ministry of Education of the People's Republic of China

Sponsor by: Hefei University of Technology

ISSN 1000-1646, CN 34-1001/N

PS-Unet: A rail surface defect detection network

Download Full Text

一种钢轨表面缺陷检测网络 PS-Unet

Journal Information

J. Hefei Univ. Tech. (Nat. Sci.), Sept.2025, Vol.48 No.9: 1192-1200, 1217

DOI: 10.3969/j.issn.1003-5060.2025.09.006

Authors

XU Jianjun $ ^{1} $, HU Xiangtao $ ^{1} $, ZHANG Yongle $ ^{1} $, LI Ziyi $ ^{1} $, ZHAN Honghui $ ^{2} $

(1. School of Electrical Engineering and Automation, Anhui University, Hefei 230601, China; 2. HUST-Wuxi Research Institute, Wuxi 214174, China)

Abstract and Keywords

Abstract: A defect segmentation model named PS-Unet based on Unet network is proposed. On the basis of Unet structure, pyramid pooling module (PPM) and scale-aware module (SAM) are developed and embedded in the deep feature extraction layer to improve the performance of the model. In addition, to address the model accuracy degradation caused by the defect-background pixel imbalance, an improved loss function is proposed to make the training process focus on the defects that are not easy to identify, and accelerate the model convergence. The experimental results show that the proposed PS-Unet network significantly improves the detection performance. Compared with Unet network, the mean intersection over union (mIOU), mean pixel accuracy (mPA) and precision (P) increase by 2.06%, 4.98%, and 3.18% respectively on the high-speed rail defect datasets, and by 4.79%, 4.35%, and 6.28% respectively on the ordinary/heavy-duty transport rail defect datasets.

Keywords: rail defects; codec; pyramid pooling; scale-aware

作者信息

许建军 $ ^{1} $，胡祥涛 $ ^{1} $，张勇乐 $ ^{1} $，李子怡 $ ^{1} $，湛红晖 $ ^{2} $

(1. 安徽大学电气工程与自动化学院，安徽合肥 230601；2. 华中科技大学无锡研究院，江苏无锡 214174)

摘要和关键词

摘要: 文章提出一种基于Unet网络的缺陷分割模型PS-Unet，该模型在Unet网络结构基础上，设计金字塔池化模块（pyramid pooling module，PPM）和尺度感知模块（scale-aware module，SAM），并将其嵌入深层特征提取层中，用于提升模型的性能；针对缺陷和背景像素比例失衡导致的模型精度降低问题，提出一种改进的损失函数，使训练过程聚焦在不易识别的缺陷上，加快模型收敛。实验证明，该文提出的PS-Unet网络显著改善了检测性能，相较Unet检测网络，平均交并比（mean intersection over union，mIOU）、平均像素精度（mean pixel accuracy，mPA）、精确率P在高速轨道缺陷数据集上分别提高2.06%、4.98%、3.18%，在普通/重型运输轨道缺陷数据集上分别提高4.79%、4.35%、6.28%。

关键词: 钢轨缺陷；编解码；金字塔池化；尺度感知

Funding

国家自然科学基金资助项目（52175210）

Previous Next

Back to Contents