Ordered Cross-Scale Interaction Network for No-Service Rail Surface Defect Segmentation

No-service rail surface defect (NRSD) segmentation plays a key role in industrial intelligent manufacturing to achieve pixel-level defect localization and ensure the quality of rails. However, the unique properties of no-service rails, such as blurred boundaries and limited semantic information, make surface defect segmentation extremely challenging. In this article, we introduce the pyramid vision transformer (PVT) to NRSD segmentation and propose a novel transformer-based ordered cross-scale interaction network (OCINet). The core of OCINet is the global-local–global strategy. Obviously, a transformer backbone is arranged to extract global features. Subsequently, three cross-scale interaction modules, including cross-scale channel interaction module (CCIM), cross-scale spatial interaction module (CSIM), and cross-scale pixel interaction module (CPIM), are used to achieve ordered channel, spatial, and pixel interactions of defect features across adjacent scales. Among them, CCIM and CSIM are used for local interaction, while CPIM is used for global interaction. These modules not only locate the defect regions but also model the pixel-by-pixel relationships between defect regions and backgrounds, generating discriminative features for defect segmentation. Comprehensive experiments and analysis on the NRSD-MN dataset, which contains 4101 NRSD images, indicate that our OCINet outperforms 19 state-of-the-art methods, achieving 68.7% in mean IoU and 78.3% in mean Dice. The code and results of our method are available at https://github.com/MathLee/OCINet