In order to accurately extract structured light stripes from weld images in the complex noise environment, we proposed a deep learning model combining semantic segmentation with object detection to detect the weld images. In the semantic segmentation branch, the model was optimized by adding parallel downsampling modules and reducing the number of convolution kernels to increase the detection speed, and the feature extraction parts of this branch and the object detection branch shared the weights. Aiming at the problem that the proportion unbalance of structured light stripes and background pixels in the weld images caused the model segmentation results to be biased towards negative samples, we introduced a Dice coefficient into the loss function to correct the model. The experimental results show that the proposed method can achieve the extraction of structured light stripes with high accuracy on the basis of ensuring real-time performance.