Advancing Welding Automation: Optimized Weld Pool Segmentation Via Hybrid Model Integration and Two-Stage Data Augmentation

This paper focuses on the weld pool segmentation task and quantitatively compares the segmentation performance of three basic model architectures. To address the issue of scarce annotated data, a two-stage data augmentation strategy was designed, combining traditional techniques with feature matching to generate diverse training samples. By innovatively integrating YOLO's real-time detection capabilities with SAM's high-precision segmentation, a fully automatic and rapid segmentation approach is achieved. For the first time, a Mamba-based model is introduced in this field, demonstrating superior computational efficiency compared to Transformers when processing long-sequence high-resolution data. This study effectively enhances the accuracy and speed of weld pool segmentation through improved data augmentation strategies and innovative model architecture integration, providing an efficient solution for welding process monitoring and holding practical value for intelligent welding quality control.