Weld seam profile identification based on visual attention mechanism in robotic thick-plate welding

This paper presents a novel vision sensor to capture weld pools and weld seam profiles simultaneously in the same frame for implementing autonomous route planning in robotic thick plate welding. A method of extracting the weld seam profile based on visual attention mechanism from weld pool background is proposed here. In this method, brightness and direction are selected as primary visual characteristics, and the captured image is deemed as the brightness feature map while the direction feature map is acquired by Gabor filtering. Then, the above two feature maps are measured individually by the different methods to better highlight corresponding characteristics, and the two measured feature maps are integrated into a comprehensive saliency map by a self-judging algorithm. To extract the seam profile, threshold segmentation is applied to the comprehensive saliency map and followed by nearest neighbor clustering. The maximum cluster is the extracted seam profile and considered as the first noticed region. Experimental results show the effectiveness of the proposed method.