Design Of TFT Screen Defect Detection System Based On Deep Learning

Nowadays,Thin Film Transistor(TFT)has been widely used in the manufacturing of liquid crystal display which is important part of electronic products.Defect detection is one of the key technologies that would affect the quality of TFT screens.In most cases,feature extraction and feature matching are two essential processes in industrial detection.Usually,designers design appropriate feature descriptors which can match the defects in images to separate and recognize defects.These feature descriptors are usually based on low-level image features,which lead to their weak generalization.For the past few years,however,object detection based on deep learning has made remarkable progress because of its outstanding feature extraction ability.They can extract more complex features from images.Nevertheless,on the one hand,the amount of unqualified products is far more less than the amount of qualified products due to the maturity of production technology and advanced management methods,resulting in the unbalanced distribution of data.It means the raw data collected from industrial production line cannot be up to the mustard of training.On the other hand,a lot of labeled data also means high cost of labor,which is hard for enterprises and factories to accept.This study propose an industrial defect detection framework based on unsupervised learning.In order to reduce the information flow during the detection and decrease the noise,an algorithm based on corner detection is designed to get region of interest.Besides,this study also propose a second-order fusion model based on Generative Adversarial Networks,which could be used to identify surface defects.This model consist of discriminator,generators and encoders.The first-order model can complete the mapping between image space and latent space quickly to achieve the task of image reconstruction.The second-order model can be used to learn the error made during the reconstruction.All of them would be used comprehensively to determine the final result.According to the experiments,the proposed method could maintain an excellent detection accuracy while the false detection rate is still at a low level with noise and the absence of real defect samples.In addition to this,the model with second-order reconstruction error is better than the First-order fusion model.