The Research On Glass Defect Recognition Method Based On Deep Learning

In the glass production line,it's not allowed for a large number of obvious defections such as bubble,knob and inclusion on the glasses.Glass defects accurately identify research is helpful to improve production technology and the quality of glass.Success or not of defect feature extraction will affect the success of the defect type recognition accuracy,but the current method of feature extraction is largely depend on experience and luck,and requires a lot of time to adjust.This thesis mainly studies the method of glass defect recognition based on deep learning,which aims to realize the diversity of defect types of accurate recognition.The main work of this thesis is as follows:1).On the basis of the analysis of the basic theory of deep learning,the structure and learning algorithm of feedforward neural network(FNN),stack autoencode network(SAE)and convolutional neural network(CNN)are studied emphatically.At the same time,sample database contains common glass defect types are established.2).This thesis studies the method of glass defect identification based on FNN,SAE and CNN combined with support vector machine(SVM)respectively and mainly studies the influence of network structure on the time and accuracy of defect recognition.Based on the comparison results,CNN is selected as the deep learning model for glass defect recognition.3).In order to overcome the shortcomings of CNN,such as network parameters adjustment difficulties,easily arriving at local minimum,CNN is improved by the combination of unsupervised learning and genetic algorithm.The method is to select a set of patches from the training images and carry out ZCA whitening process,construct an autoencoder and optimize the initial weight of the autoencoder by the genetic algorithm,then train the autoencoder to extract hidden features from the patches,which are used as theconvolutional kernel of the CNN,and finally train a SoftMax classifier on top of the convolved and sampled features for defect recognition.Simulation results show that the improved CNN not only solves these problems such as the network is easy to fall into the local minimum,the recognition accuracy of glass defects is also increased to 97.8%.