Under the background of strong noise such as arc light, smoke and a lot of spark splash, conventional laser stripe extraction method based on image processing have the shortcomings of poor flexibility and robustness. In this paper, a method of laser stripe extraction based on full convolution neural network (DB-U-Net) is proposed. The experiments show that by introducing dense residual block (DB) and attention mechanism into the backbone network of the model, the global information extraction ability of the model is improved, and the comprehensive performance AUC_PR of the model is increased from 0.891 to 0.924. On this basis, the multi-level deep supervision training mode combined with multi-layer feature cascading output is used to integrate the low-level and high-level feature information, which reduces the information loss caused by multiple up and down-sampling and deep network convolution operation. The comprehensive performance AUC_PR of the model is increased from 0.924 to 0.932. By using the proposed model network, the position error of laser stripe centerline extraction after de-noising the strong noise image is up to 0.50 pixel, which proves that the method has high detection accuracy and robustness against the strong noise interference in the welding process.