<p>The development of teach-less welding robots has gained increasing importance in modern manufacturing, as precise seam extraction plays a critical role in guaranteeing welding quality and efficiency. Image processing methods based on deep learning can accurately conduct seam extraction. However, noise interference in practical applications limits extraction accuracy, and the prevalent class imbalance in seam recognition further complicates the situation. This paper proposes a multichannel feature fusion network (MMF-NET), which aims at improving seam extraction through advanced image segmentation algorithms. First, the architecture of the U-Net network is enhanced by improving the encoder and decoder components and incorporating a multichannel feature fusion (MFF) algorithm to effectively filter out irrelevant noise and enhance feature extraction capabilities. Moreover, a mask loss function is introduced to address the common class imbalance issue in seam recognition. The experimental results demonstrate that the MFF-NET architecture not only can increase the accuracy and efficiency of seam extraction but also has the potential to significantly improve the capabilities of teach-less welding robots, thereby paving the way for advancements in automated welding processes.</p>