A novel automatic classification approach for micro-flaws on the large-aperture optics surface based on multi-light source fusion and integrated deep learning architecture

Micron-level flaws on component surface can seriously affect the optical and mechanical properties of optics, so it is significant to accurately classify and repair surface flaws. However, the small scale, multiple categories, and complex characteristics of surface flaws make the accurate classification a great challenge in the field of precision optical engineering. In this work, a novel automated classification approach is proposed for multi-classification of flaws. Images of flaws illuminated by multi-light sources are automatically acquired by classification equipment to enrich the characteristics of flaws, and a series of automated workflows are set up to accomplish flaw detection, classification, and mitigation. SingleCNN-MergeRGB, a feature extraction structure based on a single convolutional neural network (CNN) and multi-light source fusion, is constructed to extract the integrated features of flaws under different illumination. The experimental results show that SingleCNN-MergeRGB has stronger capability of feature description compared with single light source and other feature extraction networks. The integrated deep learning architecture SingleCNN-MergeRGB + SVM and prediction probability threshold adjustment strategy are proposed to improve the accuracy of flaw classification and correct failure cases. Validation results on the test set show that the proposed flaw classification method obtains an accuracy of 96.34% and a recall rate of 100%. The experiments demonstrate the effectiveness of the automatic classification equipment and methods in this work in the classification of surface flaws in large-aperture optical components.