Non-local bilateral information guided molten pool edge detection

The spatial characteristics of the molten pool are related to the welding process and can quantitatively reflect the welding quality. In response to the challenges in the detection of molten pool edges, such as strong visual interference and significant grayscale variations, a non-local bilateral information-guided method for molten pool edge detection (NLBI-MPE) is proposed. Firstly, the local features of the molten pool are enhanced based on a spatial-channel hybrid attention mechanism. Secondly, the global contextual features of the molten pool edges are captured based on multi-head self-attention. Thirdly, the bilateral features in the value-spatial domains are enhanced based on relative positional encoding. The experimental results show that the spatial-channel hybrid attention mechanism enhances the interference resistance capability of edge detection; the multi-head self-attention mechanism increases the recognition accuracy of complex molten pool boundaries through strengthening the modeling of complete molten pool structures; the relative positional encoding method elevates the semantic understanding of complex molten pool images, thereby collaboratively enhancing the capability of the model to identify both local details and global structural features. The proposed NLBI-MPE method provides a technical basis for machine vision-based quantitative non-destructive evaluation of welding quality and the dynamic adjustment of welding process parameters.