Weld-SDPose: a lightweight multi-scale framework for real-time weld seam keypoints detection

Abstract

Online visual perception for thick-plate welding requires robust weld-geometry estimation under strong specular reflections, smoke, spatter, and local occlusion, while maintaining low computational overhead for online measurement and control. To address these challenges, this paper proposes Weld-SDPose, a lightweight multi-scale framework for weld seam keypoints detection. The proposed framework is designed according to the visual characteristics of structured-light weld images, especially elongated weld geometry, local high-intensity saturation, and direction-dependent structural continuity. In the backbone, a specular-robust local attention module is developed to preserve local weld contours and suppress reflection-induced abnormal responses. A directional-illumination collaborative enhancement module is further introduced to enhance directional continuity and illumination-aware feature modulation for slender weld structures. A lightweight top-down multi-scale fusion neck is used to aggregate semantic and spatial details with low computational cost. In addition, a structure-prior guided pose head and a prior-guided geometric consistency loss are introduced during training to improve the structural consistency of predicted keypoints without increasing inference overhead. Experimental results show that Weld-SDPose achieves 99.2% mAP@0.5 and 98.8% mAP@0.5:0.95 with only 0.92 M parameters and a 2.3 MB model size. The model reaches a single-frame latency of 6.0 ms on the RTX A6000 GPU and 57.0 ms using a CPU-only platform, indicating favorable inference efficiency and lightweight deployment potential under the tested hardware conditions.