Abstract: Aimed at the weld vision detection problem of arc welding robot in CCD directly photographed image, an algorithm for the extraction of special weld seam track with large discontinuity, sharp turning point and more interference was proposed. The algorithm firstly use wavelet transform to smooth the weld edge line to get the initial weld track that is very close to the final weld, and then use Snake model and genetic algorithm to get accurate matching and extraction of special welds. Experimental results demonstrate that the algorithm can improve the precision of special weld track recognition and detection, and has strong adaptability, and can also provide some technical reference for further improving the intelligent vision detection of arc welding robot.