EWSeg: A geometry-aware efficient semantic segmentation network for structured-light weld seam images

Abstract Line-structured light (LSL) vision sensors are widely used in industrial automation for precise weld seam measurement and robotic guidance. However, in real-world welding environments, intense arc light, metal spatter, and complex background noise severely degrade the quality of laser stripe imaging, leading to measurement errors. This paper proposes EWSeg, an efficient geometry-aware semantic segmentation network for structured-light weld seam images. The network uses a lightweight R-MambaOut backbone with coordinate encoding and multi-branch direction-scale convolution to improve spatial geometric representation at low computational cost. To mitigate spatial information loss during feature extraction, we introduce the Cross-Scale Geometry Fusion (CSGF) module, which enhances the continuity of weak-texture boundaries through deformable alignment. The decoder further combines Dense Atrous-Strip Pooling Context Module (DS-ASPCM) and Global Channel-Spatial Attention (GCSA) to capture multi-scale context, anisotropic stripe structure and target-relevant saliency. On the self-built dataset, EWSeg achieved 93.36% mIoU, 96.78% mPA and 61.57 FPS. At the 2D segmentation stage, EWSeg achieved a boundary IoU of 81.8% and a centerline error of 1.50 pixels, indicating improved geometric fidelity and centerline localization. Cross-scenario experiments showed that this performance remained stable across sensor configurations and lighting conditions, supporting its use as a reliable basis for downstream structured-light measurement.