Abstract: The level set image segmentation method (C-V method) was proposed by Chan and Vese based on Mumford-Shah model.Its limitation was analyzed and the improved method was provided according to the character of welding environment.It has three mainly limitation, that is, the image must have distinct characters; the target to be divided can not be too many and dispread;it coss excessive time.The improved algorithms include modification of strengthened character model, multi-scale rapid calculation and suppression of global characters.The comparison test and lab experiment in real welding environment were given.The results prove the improved algorithm can extract feature accurately and adapt to environment changes.The improved algorithms also provide a good way to deal with image character extracted for complex images.