Printed Circuit Board (PCB) is used as a carrier for electrical interconnection of electronic components. With the rapid development and wide adhibition of integrated circuits, the volume of electronic devices is getting smaller and smaller, and the density and difficulty of circuit wiring are also getting greater and greater. In industrial production and manufacturing, hundreds of components are often glued to a PCB board, which requires high precision. Deep learning has been widely used in PCB defect detection for its excellent performance. However, in practical applications, methods based on deep learning often suffer from overfitting problems due to lack of sufficient training data. At the same time, these methods still have challenges in detecting these small size and irregular shape defects. In order to solve these problems at the same time, this paper develops a new small sample learning mode based on transfer learning with Unet as the basic framework and small sample meta-learning as the basis method. Meanwhile, in order to capture the defects of different scales and rules, We propose a general fet-shot PCB Semantic Segmentation Based on Transfer Learningand Multi-Scale Fusion (FSPSS) to complete PCB image defect segmentation. Experiments on PCB defect datasets demonstrate that under different lenses (k=1,2,3,5,10), our model outperforms the most advanced methods and has better generalization ability.