Abstract: In order to reduce the influence of weld shape and color diversity on segmentation accuracy in laser weld semantic segmentation, an image semantic segmentation method based on attention mechanism was used to extract weld. The image in the weld was converted from RGB(red, green, blue) to HSV(hue, saturation, value) color space, and the weld surface color was recognized in HSV. The effects of three kinds of welds on region segmentation and color recognition were analyzed. The results show that the average pixel accuracy of the weld segmentation region is about 91.2%, and the segmentation effect of the attention U-Net model with attention mechanism is better. The results of automatic identification of weld surface color meet production requirements, and have broad application prospects in industrial production.