Abstract: Due to the influence of complex time-varying interference and variations in process conditions during the welding process, the boundary characteristics of the molten pool are easy to be blurred, and the scale information is complex and changeable, which poses significant challenges to the accuracy recognition and the robust segmentation of the molten pool. In this paper, a molten pool image recognition method combining attention mechanism and transfer learning is proposed. Firstly, the residual block(RB) is added to the UNet down-sampling process to extract multi-scale low-level features, and the coordinate attention block(CAB) is introduced in the down-sampling and up-sampling processes to improve the feature weight of the effective region. Secondly, the pre-trained deep convolutional neural network in Pascal VOC2012 is transferred to the UNet network to realize feature transfer and parameter sharing, so as to alleviate the over-reliance of training effect on datasets. The TL-RCUNet network proposed in this paper has achieved good recognition results on the untrained MAG and TIG cross-process datasets. The mean intersection over union(MIoU) reaches 96.21% and 79.55%, respectively, which is about 15% and 25% higher than the classical semantic segmentation network. The model provides a feasible solution to the problem that existing semantic segmentation methods rely on a large number of training samples and pixel-level annotations based on expert experience.