Enhanced transformer method coupled with transfer learning for surface defect segmentation of myopia control spectacle lenses.

Myopia requires visual correction. The complications associated with myopia affect a large population of schoolchildren around the world. Nanostructured myopia control spectacle lenses (NMCSLs) containing nano surface features are commonly used as a non-invasive approach for slowing down the progression of myopia. However, the effective segmentation of surface defects generated in the precision manufacturing of the NMCSL heavily relies on highly efficient and effective defect detection and characterization methods. As a result, this paper presents an enhanced transformer method coupled with the transfer learning (E2Trans) method, which combines the powerful feature extraction abilities of the transformer and the knowledge re-usage abilities of transfer learning to realize high-efficiency and high-accuracy defect segmentation. To further improve the segmentation performance, two auxiliary decoders are added to adjust the training loss. To validate the model's performance, a lens defect dataset is built, and a series of experiments are conducted. The results show that our proposed model can segment five lens defects, including notches, black spots, bubbles, fibers, and scratches with high segmentation accuracy and speed. In addition, a detection system is developed for real-time lens defect detection.