RBA-Net: A Rail Surface Defect Segmentation Network with Region and Boundary Perception

Based on this idea, we propose a rail surface defect segmentation network called RBA-Net, which utilizes region and boundary perception. The network follows an encoder-decoder structure, with each stage of the encoder connected to the corresponding part of the decoder through three modules.

The first module, called the Feature Pyramid Edge (FPE) module, aims to capture multi-granularity edge information of defects. The second module, the Multi-Task Learning (MTL) module, oversees defect region segmentation and boundary detection. To fully leverage the information between these two tasks, we introduce an Information Interaction Layer (IIL) that enhances the performance of defect segmentation.

Lastly, the Cross-layer Fusion (CLF) module is designed to further capture context by selectively aggregating multi-layer features from the encoder, resulting in more comprehensive spatial information. By cascading the FPE, MTL, and CLF modules, the proposed RBA-Net effectively incorporates rich context information and fine-grained features, promoting accurate segmentation of rail surface defects.