Laser vision seam tracking enhances robotic welding by enabling external information acquisition, thus improving the overall intelligence of the welding process. However, camera images captured during welding often suffer from distortion due to strong noises, including arcs, splashes, and smoke, which adversely affect the accuracy and robustness of feature point detection. To mitigate these issues, we propose a feature point extraction algorithm tailored for weld images, utilizing an improved Deeplabv3+ semantic segmentation network combined with EfficientDet. By replacing Deeplabv3+’s backbone with MobileNetV2, we enhance prediction efficiency. The DenseASPP structure and attention mechanism are implemented to focus on laser stripe edge extraction, resulting in cleaner laser stripe images and minimizing noise interference. Subsequently, EfficientDet extracts feature point positions from these cleaned images. Experimental results demonstrate that, across four typical weld types, the average feature point extraction error is maintained below 1 pixel, with over 99% of errors falling below 3 pixels, indicating both high detection accuracy and reliability.