LCD Screen Defect Classification System Based On Transfer Learning

With the popularity of smart phones,LCD(Liquid Crystal Display)screen classification has become a focus in the field of industrial vision.Industrial screen defect recognition scene is a small sample scene,and the positive and negative samples are not balanced.At present,the identification of screen defects mainly relies on manual quality inspection and traditional machine vision detection methods.Manual quality inspection has the problems of low efficiency,inconsistent standards and high cost.Compared with manual quality inspection,the traditional machine learning method based on digital image processing has an improvement in speed,but the accuracy still needs to be improved.Ordinary deep learning algorithm needs a lot of data,and it is easy to over fit in our scenario.Therefore,in order to solve the above problems,this thesis aims to improve the accuracy and efficiency of LCD screen classification by using transfer learning combined with image processing algorithm and deep neural network model.The main contents of this thesis are as follows:Firstly,this thesis designs an image preprocessing module.This module can get ROI of the original image to eliminate the bad influence of black background on the classification accuracy,and also greatly reduce the memory consumption.Secondly,this thesis designs an auto generation module with the help of spatial and morphological changes.The generation module is based on the analysis results of defect features to simulate.The function of the generation module is to simulate real defective screen samples and generate artificial ones with similar characteristics to expand the industrial image set.The generation module is closely related to the accuracy,and mixing proportion of real defects and generated ones also affects the performance of the system.This thesis finds the best mixing proportion.Finally,the thesis proposes an industrial screen classification system based on the above preprocessing module,generation module and transfer learning idea,which uses a variety of technologies to make up for the problem of too few negative samples of real screen defects in the industrial field.This thesis compares our method with a variety of models,and obtains the conclusion of transfer learning is more conducive to improve the accuracy of our system.At the end of the paper,we summarize the research content of this paper,and also provide a reference for the follow-up research in the industrial field.