SABE-YOLO: Structure-Aware and Boundary-Enhanced YOLO for Weld Seam Instance Segmentation.

Area of Science:

• Computer Vision
• Robotics
• Materials Science

Background:

• Automated welding systems require precise weld seam recognition for path planning and quality control.
• Current industrial vision approaches for weld seam instance segmentation face challenges with complex geometries and blurred edges.
• Existing models struggle with accurate boundary perception and structural representation of elongated weld seams.

Purpose of the Study:

• To develop an advanced weld seam instance segmentation method addressing limitations in boundary perception and structural representation.
• To enhance the accuracy and efficiency of visual perception for complex weld seam structures in automated welding.
• To introduce a novel approach that improves localization accuracy and feature extraction for weld seams.

Main Methods:

• Proposed a novel structure-aware and boundary-enhanced YOLO (SABE-YOLO) model for weld seam instance segmentation.
• Introduced a Structure-Aware Fusion Module (SAFM) utilizing strip pooling attention and element-wise multiplicative fusion for enhanced structural features.
• Developed a C2f-based Boundary-Enhanced Aggregation Module (C2f-BEAM) for improved edge sensitivity and multi-scale feature aggregation.
• Implemented an Inner-MPDIoU loss function to boost localization accuracy.

Main Results:

• SABE-YOLO achieved a 46.3% AP(50-95) metric, outperforming YOLOv8n-Seg by 3 percentage points.
• The model demonstrated efficient computational performance with 18.3 GFLOPs and 6.6M parameters.
• Achieved a high inference speed of 127 FPS, indicating a favorable balance between accuracy and efficiency.
• Experimental results validated the effectiveness on a self-built weld seam image dataset.

Conclusions:

• The proposed SABE-YOLO provides an effective solution for high-precision visual perception of complex weld seam structures.
• The method shows strong potential for practical industrial applications in automated welding.
• SABE-YOLO offers a superior trade-off between segmentation accuracy and computational efficiency compared to existing methods.