Deep Learning Based Optical Component Surface Defect Detection Technology Research

With the rapid development of optical technology,optical components have been widely applied in fields such as aerospace,defense,military,medical,and communication.However,during the manufacturing process of optical components,defects such as spots and scratches may occur on their surfaces due to processing issues.These defects severely impact the stability of optical systems composed of these components and even affect the imaging quality of the entire optical system.Traditional defect detection algorithms have drawbacks such as complexity and low accuracy,and they are no longer able to meet the requirements of high-precision detection for optical components.Therefore,this study focuses on the accurate measurement of surface defects on optical components,and the main research contents are as follows:Through the analysis of surface defect characteristics on optical components,a defect detection method is proposed that combines deep learning image extraction algorithms with darkfield microscopy imaging.The dark-field images of defects serve as the data source for the defect extraction network.A deep learning network model suitable for surface defect datasets of optical components is designed and constructed.By training and testing the model,defect binary images are obtained through neural network-based extraction.Further analysis is conducted on the defect binary images to complete the measurement of the two-dimensional dimensions of surface defects on optical components.Establishment of the surface defect dataset for optical components.Based on dark-field imaging theory,a dark-field microscopy detection device is constructed to capture dark-field images of surface defects on optical components.The Gaussian bilateral filtering algorithm is applied to remove noise interference from the images based on the characteristics of defect images.Then,the images are annotated,and to address the problem of insufficient sample data,data augmentation techniques such as image flipping,rotation,and translation are used to expand the dataset.This completes the creation of a dataset of dark-field microscopic images of surface defects on optical components,providing data support for the deep learning defect extraction model.Research on accurate extraction of surface defects on optical components using the improved U~2-Net network.To enhance the accuracy of defect extraction by neural networks,attention modules and residual network modules are introduced to the U~2-Net network.These modules effectively address the issue of partial feature loss of surface defects on optical components during network training and solve the problem of gradient disappearance that arises with deeper networks.This reduces the training difficulty of the network model and enhances the effectiveness of surface defect extraction.While maintaining extraction accuracy,the traditional3x3 convolutions in the downsampling layers of the original network are replaced with depthwise separable convolutions to reduce computation and parameter requirements,resulting in a reduction of training time by nearly 5 hours.From the training and testing results of the model,the improved network model demonstrates faster convergence speed,with pixel-level defect extraction accuracy reaching 95.7%,Dice coefficient reaching 91.3%,and Intersection over Union(IOU)reaching 91%.This indicates that the improved U~2-Net network can accurately extract surface defects on optical components.Experimental research on surface defect detection of optical components.Based on the improved U~2-Net network,binary defect images are extracted from the surfaces of optical components,and the two-dimensional dimensions of scratches and spots are quantified.Firstly,skeletonization processing is applied to the defect binary images,and through analysis of the defect skeletons,the two-dimensional dimensions of surface defects on optical components are calculated using defect size calculation methods.Then,the defect dimension calculation results are compared with the results obtained from a white light interferometer for detection,and the detection results are found to be consistent,demonstrating that the proposed method can achieve fast and accurate detection of surface defects on optical components.