Abstract: Objective Welding plays a crucial role in the energy and petrochemical industries. The arc welding process involves complex physical information, and the dynamic behavior of the molten pool directly affects weld formation, as well as determines the stability of the welding process and the welding quality. Currently, numerous methods based on visual and deep learning models have emerged in the field of welding molten pool dynamic monitoring. However, these methods require extensive labeled data to train network models, consuming significant human resources. Methods To address these issues, this study proposes a visual monitoring method for the welding molten pool based on semi-supervised semantic segmentation. Additionally, to tackle the challenge of the variable shapes of welding molten pool image features, this work introduces a novel multi-shape feature information enhancement module that combines strip convolution, dilated convolution, and traditional convolution in a parallel configuration to improve the accuracy of the semantic segmentation network. Results Experiments conducted on a self-built dataset show that the proposed method achieves 48.5% m_IoU with 850 labeled data samples. Conclusion This work can be applied to quality analysis of the welding molten pool and holds significant importance for achieving fully automated management of the welding molten pool.