In laser vision-based welding tracking, extracting laser line features from welding images to accurately locate the welding position encounters challenges, especially when strong arc light interference causes the laser line to be blocked. Existing methods often struggle to achieve satisfactory results in the face of continuous and robust noise interference. Consequently, we introduce a novel two-stage Arc Light Elimination Network (ALENet). Specifically, within the coarse elimination stage of ALENet, the network learns the distribution information of interference from the potential feature space, enabling effective filtering to obtain preliminary results. In the subsequent finetuning stage, leveraging the current polarization image and the prior knowledge of laser line image provides enriched spatial distribution information and texture structure features to obtain the final laser stripe image. Furthermore, we collected and created a Welding Arc Light Elimination Dataset (WALED) for rigorous experimental validation. The experimental results show that ALENet has excellent performance on WALED and can restore clear laser streak images even under noise occlusion, such as strong arc light interference.