Open Access

ARTICLE

Rail Surface Defect Detection Based on Improved UPerNet and Connected Component Analysis

by Yongzhi Min^1,2,*, Jiafeng Li³, Yaxing Li¹

1 School of Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou, 730070, China
2 Gansu Provincial Engineering Research Center for Artificial Intelligence and Graphics & Image, Lanzhou Jiaotong University, Lanzhou, 730070, China
3 State Grid Gansu Electric Power, Baiyin Power Supply Company, Baiyin, 730900, China

* Corresponding Author: Yongzhi Min. Email:

Computers, Materials & Continua 2023, 77(1), 941-962. https://doi.org/10.32604/cmc.2023.041182

Received 13 April 2023; Accepted 14 August 2023; Issue published 31 October 2023

Abstract

To guarantee the safety of railway operations, the swift detection of rail surface defects becomes imperative. Traditional methods of manual inspection and conventional nondestructive testing prove inefficient, especially when scaling to extensive railway networks. Moreover, the unpredictable and intricate nature of defect edge shapes further complicates detection efforts. Addressing these challenges, this paper introduces an enhanced Unified Perceptual Parsing for Scene Understanding Network (UPerNet) tailored for rail surface defect detection. Notably, the Swin Transformer Tiny version (Swin-T) network, underpinned by the Transformer architecture, is employed for adept feature extraction. This approach capitalizes on the global information present in the image and sidesteps the issue of inductive preference. The model’s efficiency is further amplified by the window-based self-attention, which minimizes the model’s parameter count. We implement the cross-GPU synchronized batch normalization (SyncBN) for gradient optimization and integrate the Lovász-hinge loss function to leverage pixel dependency relationships. Experimental evaluations underscore the efficacy of our improved UPerNet, with results demonstrating Pixel Accuracy (PA) scores of 91.39% and 93.35%, Intersection over Union (IoU) values of 83.69% and 87.58%, Dice Coefficients of 91.12% and 93.38%, and Precision metrics of 90.85% and 93.41% across two distinct datasets. An increment in detection accuracy was discernible. For further practical applicability, we deploy semantic segmentation of rail surface defects, leveraging connected component processing techniques to distinguish varied defects within the same frame. By computing the actual defect length and area, our deep learning methodology presents results that offer intuitive insights for railway maintenance professionals.

---

Keywords

---