Defect Recognition Of TFT Panel Gate Layer Based On Convolution Neural Network

TFT panel manufacturing is prone to a variety of defects,classification and detection work is particularly important.At present,the traditional artificial vision detection method is still used in detection and classification,which is susceptible to subjective factors and external environmental impact,and it is difficult to ensure product quality and low efficiency.Therefore,there is an urgent need to study a stable and efficient defect recognition method.This paper takes Gate layer of TFT panel as the research object,and develops an automatic recognition algorithm based on deep learning.It is found that the defect map of Gate layer after stripping is the most suitable research object for KOI machine under VPD.After obtaining the image data,the image is calibrated according to the classification criteria,and then the image preprocessing,including gray scale,scaling,normalization and other steps.Because of the small data set,data expansion is needed.We try to use image geometry transformation to expand data in three ways: against generating network and artificial forgery.The paper studies and implements the Gate Layer Defect Recognition Deep Learning Network.Firstly,traditional machine learning algorithms,such as support vector machine(SVM),logical regression,nearest neighbor algorithm(KNN)and random forest,are explored and used for defect recognition.Secondly,convolutional neural network is emphatically implemented for defect classification.A shallow convolution neural network VTest with 8 layers of convolution is designed and its parameters are optimized,including convolution,pooling,peripheral filling,regularization,activation function and loss function;VGG16/Inception_v3/Resnet50/Inception_resnet_v2 network is used for ab initio training and migration learning;Inception_resnet_v2 structure is modified as a basic module to build a multi-layer network for training.DCGAN is constructed for image expansion.The experimental results show that:(1)compared with the traditional machine learning method,the accuracy of convolutional neural network is 10% higher;and(2)the accuracy of convolutional neural network is 10% higher after the expansion of training data.(3)The results of ab initio training on Tensorflow platform show that the accuracy rate of Inception_v3/Resnet50 structure is 77%,which is increased by 17% again.The accuracy of ab initio training on Inception_resnet_v2 network is 88%.The average training time from the beginning is up to 10 hours.(4)The network structures on Keras platform use transfer learning training method.The accuracy of VGG16/Inception_v3/Resnet50 is 84%,and that of Inception_resnet_v2 is 90%.The average training time of transfer learning is 35 minutes.Keras used Inception_resnet_v2 in the four-tier cascaded network,the accuracy of multi-classification reached 93%.The whole experiment shows that the multi-level training is suitable for the defect classification of Gate layer in TFT panel,and has a high accuracy rate to the specified design index.The convolution neural network Inception_resnet_v2 has superiority over other algorithms.Further optimization can be applied to replace manual repair judgment and assistant manual work defect classification in actual production.