Mask decoupling framework for rail surface defect pixel-level detection

Liu, Jianwei

doi:10.1016/j.aei.2026.104875

Mask decoupling framework for rail surface defect pixel-level detection

2026-06-01

0

PRE

AI

DOI：10.1016/j.aei.2026.104875

Original

Original request for help

Save

Abstract

En 中文

Rail surface defect segmentation is a critical step in assessing the severity of rail diseases, yet existing methods struggle with the imbalanced edge pixel distributions that degrade edge prediction. Aiming at this issue, a novel mask decoupling framework (MDF) that integrates the mask decoupling and a dual-encoder-based feature interaction network (FIN) is proposed. Firstly, original mask is decomposed into edgelet mask that focuses on regions around edges and body mask that concentrates on defect center regions. Different from traditional edge supervision, edgelet mask works better as it involves much more pixels at the edges. Then, FIN is constructed to adapt to edgelet mask and body mask, respectively. In which, the embedded feature attention module (FAM) is used to interact the features from two complementary encoders. While the cross feature fusion (CFF) module is responsible to aggregate the multi-level features from encoders to predict the segmentation mask. Comprehensive experiments on the expanded rail surface defect dataset (RSDD) demonstrate MDF achieved excellent segmentation performance for defect regions and edges, and outperforms state-of-the-arts across different evaluation metrics. The codes will be available at https://github.com/qiuyuan666/MDF-for-Rail-Defect-Detecti on.

Keywords:

Railway transportation

Rail surface defect segmentation

Mask decoupling

Feature interaction

Mask decoupling framework for rail surface defect pixel-level detection

Abstract

Journal

Organization

Cited Papers