Research On The Classification And Detection Technology Of Crystal Defects Based On Deep Learning

In high-power laser driver systems,the crystal is one of the indispensable optical components,mainly used as a Pockels cell(photoelectric switch)in the main amplifier system to modulate the incident light field and as a terminal optical focusing system used to generate ultraviolet light.Since the crystal size in the system is usually large and the number is large,the inspection and maintenance task for the component is relatively heavy.According to statistics,there are 576 KDP crystals of 410 mm diameter used in the national ignition device.In the production and processing,these crystals often cause different degrees of unpredictable light field modulation,wavefront distortion,which even may damage the entire drive system.Therefore,we should and need to conduct research on classification detection technology for crystal component defects.In addition to the human eye judgment,most of the existing defect detection technologies are developed around hardware devices.The algorithm is mainly based on traditional image processing,mainly for feature extraction of different types of defects,and less on the detection and identification of defect classification and positioning.Moreover,because the shape of crystal defects is complex,traditional image processing algorithms are difficult to extract key and objective features,which greatly limits the intelligent development of crystal defect detection technology and the application of high-power laser drivers.Aiming at the rich and varied features of surface defects of crystal components,this paper proposes a deep learning model for classification of crystal defects.A crystal defect detection model based on candidate region generation module and feature extraction module(ZF net)is constructed.At the same time,a database of 1800 crystal defects covering hydrolysis,pits and cracks was constructed.Subsequently,this paper develops the model based on the Tensorflow framework and the Python locale,and tests and verifies the ZF net-based detection model.The parameters are driven by the characteristics of the crystal defect dataset,and the loss function value of the model is controlled below2.5*10-5.The results show that the detection and recognition ability of crystal defects is strong,the recognition accuracy of different defect categories can reach more than 86%,the false detection rate is less than 59%,and the detection results have high reliability.In terms of efficiency,the detection time of a picture is about 100 milliseconds,which can cope with scenes with low real-time requirements,thus replacing manual detection and traditional image processing detection,and improving the efficiency of traditional detection while ensuring accuracy.In addition,based on the database size constructed in this paper,this study also uses the VGG16-based detection model to compare with the ZF net-based detection model.The results show that the detection model based on VGG16 has higher accuracy and its average detection accuracy is about 2.5 percentage points higher than the ZF model on the verification set.In terms of efficiency,VGG16 is 1/3 of ZF,but the time is still controlled on the order of 100 milliseconds.In practical applications,it is necessary to comprehensively consider the accuracy and detection efficiency to meet different detection requirements.In summary,the work of this paper can provide a new idea and feasible solution for crystal detection,which lays a good foundation for the further application of deep learning on crystal defects and the classification of full-field quantitative scanning results of large-diameter crystals.By comparing the results of crystal defect detection under different scale convolution networks,it can also provide reference for subsequent optimization algorithms.