<p>In automated welding, accurate weld seam tracking is essential for improving manufacturing quality and efficiency. Vision-based laser stripe centerline extraction is a key technique for seam tracking, but intense arc light, spatter, and reflections in welding environments often degrade image quality and reduce extraction robustness. To address this problem, this study proposes an improved DeepLab V3 + semantic segmentation model for robust laser stripe extraction in high-noise welding images. MobileNetV3 is adopted as the backbone to improve inference efficiency. A context-anchored attention module is introduced into the encoder to enhance the perception of elongated laser stripe structures and suppress noise responses. In the decoder, DySample dynamic upsampling is used to improve edge and detail reconstruction. In addition, multiscale feature fusion is employed, and a combined Focal Loss and Dice Loss function is used to alleviate foreground–background imbalance. On the held-out test set, the proposed method achieved 94.96% mean intersection over union, 97.96% mean pixel accuracy, 93.93% precision, 95.92% recall, 94.91% F1/Dice score, and an inference speed of 53 frames per second. Compared with selected general semantic segmentation baseline models, the proposed method provides a favorable balance among segmentation accuracy, inference speed, and model size and offers effective preprocessing for downstream laser stripe extraction and weld feature-point localization under the tested high-noise welding conditions.</p>