Research On Surface Defect Detection And Classification Algorithm Of Sapphire Based On Deep Learning

Sapphire has become a tremendously important substrate material in the field of semiconductors,especially LED field,due to its unique lattice structure and superior properties in physics,optics,chemistry and mechanics.In the process of making sapphire substrate,a series of mechanical and chemical processes are required,so some surface defects are inevitable.Surface defects will affect the yield of subsequent epitaxial layers and related devices.How to quickly and accurately locate the surface defect information and effectively classify the surface defect types is of great importance for the quality detection of substrate surface and the study of the influence of process parameters on surface quality.In this paper,sapphire substrate surface defects are taken as the research object.In order to meet the needs of rapid,high accuracy and intelligent detection and classification development,the deep learning method is adopted to focus on the research of sapphire substrate surface defects detection and classification algorithm.Specific research contents are as follows:(1)Image data set preparation.Firstly,the automatic image acquisition system of line scan is used to collect the image of defects on the surface of sapphire substrate.Then the defect image was expanded by image inversion,rotation and data enhancement technologies such as GAN.Then,LabelImg labeling software was used to manually label the defect images after expansion.Finally,the labeled defect images were divided into training set,verification set and test set.(2)Training and testing of the target detection algorithm model.Firstly,network parameters(learning rate,training steps,etc.)in the training process of Faster RCNN,SSD,and YOLOV3 detection algorithm model were developed.Then it was trained and tested on data set A respectively,and the predicted results are compared and analyzed.Finally,in order to further test the performance of YOLOV3 algorithm model under more defect types,it was trained and tested on data B.The experiments showed that the detection speed,detection and classification accuracy of YOLOV3 on data set A was higher than that of Faster RCNN and SSD.Meanwhile,the mAP of YOLOV3 on data B is 95.96%.(3)Parameters optimization of YOLOV3 detection algorithm model.Due to the cases of missing,mischecking and repeated detection in YOLOV3 testing on data B,so as to improve the above situation,the YOLOV3 detection algorithm model was optimized in three aspects: learning rate,NMS threshold value and anchor box size.Through the comparison of results before and after the optimization experiment,it can be seen that the optimized YOLOV3 algorithm model not only improves the cases of missed detection,misdetection and repeated detection,but also reaches 97.41% mAP in the end.According to the above experiments,the surface defects of sapphire substrate were detected and classified,and the accuracy of detection and classification of YOLOV3 on data set B was optimized.This paper provides a rapid and high accuracy intelligent detection method for surface defect detection and classification of sapphire substrate.