Welding seam recognition method based on DeepLabV3+ with efficient channel attention integration

Objective Welding seam recognition plays a crucial role in material processing and welding process. To address the influence of welding seam stripes segmentation accuracy of laser welding due to noise such as arc light and smoke in complex environments, an improved DeepLabV3+ welding seam recognition method is proposed, which incorporates the ECA module to enhance robustness of the model. Methods The ECA attention mechanism is introduced before the feature fusion in decoder of the model to achieve weighted feature fusion. Subsequently, a combination of cross-entropy loss, dice loss and focal loss is used to further improve accuracy and robustness of the model. Results Experimental results show that the proposed algorithm achieves excellent segmentation performance in actual welding environments, with an mPA value of 95% and an mIoU value of 89%. It effectively extracts and recognizes welding seam features. Conclusion Experimental validation for weld seams of laser welding recognition in complex environments demonstrates that the improved model significantly enhances the performance of welding image recognition, showing strong potential for practical application.