To meet the requirements for conciseness and high accuracy in algorithms for detecting surface defects on high-precision optical components，with scratches and pockmarks as the research targets，a method for detecting surface defects on optical components based on an improved U2-Net is proposed. First，the network’s dataset is constructed using defect information from the surface of optical components. The improved U2-Net network is used for real-time training and testing on the surface defect dataset of the components to be applied，and ultimately，the new method is compared and analyzed with the previous U2-Net network. Experimental results show that the new network model has reached 95.7% in accuracy，91.3% in similarity coefficient，91.2% in intersection over union，and 91.3% in recall rate for key performance indicators. This technology is capable of resisting interference from noise points，used for detecting and displaying scratches and pockmarks defects in images，while improving the accuracy of defect segmentation and the accuracy of detection，achieving effective identification and segmentation of surface defects on optical components.