Research On The Detection Methods Of Glass Defect Based On Convolutional Neural Network

With the rapid development of glass production technology,people have higher and higher requirements for glass quality.However,due to the influence of materials,environment,temperature and other factors,glass products will produce various types of stripe defects,which will affect the transparency and quality of glass.Therefore,the realization of rapid detection of glass stripe defects is of great significance to improve the glass production process and glass quality.This paper focuses on the research of the convolution neural network for glass defect detection,and the main work is as follows:Firstly,aiming at the problem of low efficiency and time-consuming in visual inspection of glass defects,this paper constructs the glass defect detection model of Yolov3,and realizes the rapid and efficient detection of glass defects.Inputting the glass image into Yolov3’s backbone network Darknet-53,and the defects of glass image are predicted by three prediction feature maps,each prediction feature map uses three scales to predict.The darknet-53 network of this algorithm uses convolution layer instead of pooling layer,so the detection effect is improved,and the reduction of convolution kernel makes the image detection speed significantly improved.Through the test of the model,the accuracy of detection model is about75%.Secondly,in order to further improve the detection accuracy,this paper constructs the glass defect detection model of Faster R-CNN.Inputting the glass image into the backbone network of Faster R-CNN to get the feature maps,the feature maps obtain the anchor boxes through RPN,and then maps the anchor boxes into the feature maps to get the feature matrices,then the feature matrices are scaled to 7x7 by ROI pooling.After a series of full connection,the target classification probability and border regression parameters are obtained,the final position and probability of the prediction target are obtained by adjusting the position of the prediction box through the border regression parameters.The accuracy of glass defect detection model based on Faster R-CNN is about 77%.In order to meet the needs of real-time detection,yolov3 algorithm with more balanced detection speed and accuracy is selected to realize the detection model of the detection system.The system realizes the automatic recognition and classification of glass image defects,and gives the proportion of three layers,at the same time,it meets the functions of manual correction,classification and marking.The system test results show that the average detection accuracy of the system reaches 78.5%,which can realize the intelligent,accurate and rapid detection of glass defects,and provide an effective way for the digital and intelligent transformation and upgrading of glass enterprises.