Abstract: In the vision detection methods for shooting the image directly with CCD of arc welding robot, This paper proposes a practical recognition algorithm for complex weld images with black note marks, scratches, and cross interferences with welds, and puts the emphasis on several key issues of this algorithm, such as the skeleton thinning, splitting, merging and pruning of edge chains, and the judgments of location relations based on prior knowledge, etc. The experimental results demonstrate that the algorithm can effectively remove the black notes, scratches and cross interferences and can automatically extract the weld track in complex weld images. The algorithm has strong adaptability and expandability, and can be applied to other complex images with a slight improvement，such as images with arc light or structured light, etc. The algorithm can provide some technical reference for further improving the intelligent vision detection of arc welding robot.