To address low accuracy and unstable quality in traditional mitre elbow weld recognition, this study aims to develop a specialized model with attention mechanisms and optimized loss functions for reliable segmentation.
A tungsten electrode height detection method is proposed. A weld recognition model MENet based on DeepLabV3+ is developed, introducing ECA attention for weighted feature fusion in the decoder. Dice loss and focal loss are added to the loss function to correct background bias, enhancing segmentation accuracy and robustness.
The proposed method demonstrates excellent performance in mitre elbow weld recognition, effectively resisting interference from workpiece reflection, noise and severe surface scratches, improving segmentation accuracy and robustness.
This research innovatively integrates the ECA attention mechanism into DeepLabV3+ for weld feature fusion and combines dice loss with focal loss, providing an effective solution for complex pipe fitting welding scenarios like mitre elbow welding, with potential for real-time monitoring and adaptive control in automated welding systems.
