Detection Of Appearance Defects And Image Generation Algorithm Of Polarizer

Polarizer is one of the imaging components of LCD,which affects the imaging effect of LCD.Therefore,the inspection requirements of polarizer surface quality are very strict.However,in the actual production line of automatic optical inspection,it is easy to miss the fine transparent defects of polarizers.In view of these tiny transparent defects,it is imperative to develop an automatic detection system with high speed,high precision and high reliability,which can also promote the development of polarizer manufacturers and related industries.In this paper,the method of saturation imaging of structured light is used to detect the extremely slight transparent aesthetic defects of polarizer.The effects of three parameters,i.e.the width ratio of black and white stripes of structured light,object distance and exposure parameters of camera,on the defect saturation imaging are studied in detail.In order to detect extremely slight transparent aesthetic defects of polarizer by using saturation imaging of structured light,the width ratio of black and white stripes of structured light can be set in the range of 1.0 ? 1.45.The object distance of camera should be controlled in the range of 50 ?200mm and the exposure value of camera exposure parameters should be controlled in the range of 3 ? 4.At the same time,a prototype is developed to detect the aesthetic defects of polarizers by using saturation imaging of structured light.Using Solid Works to build the prototype model,the 3D dimension of the prototype is 2000 × 1100 × 700 mm.The structure of loading platform and defect marking device is designed.The effective stroke of feeding mechanism is determined to be 1500 mm.The main structure of the prototype is built and testing program of the prototype based on Lab VIEW is compiled.In this paper,convolution neural network is used to detect and classify defects in polarizer images.In order to detect whether the polarizer image has defects or not,firstly,2160 groups of polarizer image without defects were collected,and the normalized histogram characteristics of polarizer image without defects which can be used to make data set were analyzed.First,the peak value of high gray value is greater than the peak value of low gray value.Second,the gray value of the peak value of low gray value is less than 50,and the gray value of the peak value of high gray value is 255.644 flawless polarizer images and 400 flawed polarizer images are selected as image data sets.By using the migration learning method,resnet50 is used to pre train the polarizer image binary classification network.The accuracy of the verification set of the binary classification network is 98.36%,and the accuracy of the test set is 97.00%.It shows that the trained binary classification network can be used to detect whether the polarizer image has defects or not.Aiming at the problem of multi classification of polarizer defect images,there are 400 real polarizer defect images.The generalization ability of multi classification model based on Res Net50 pre training is poor.The accuracy of test set is only 77.50%,far lower than 90.00%of training set.In order to solve the problem of insufficient number of polarizer defect images,a polarizer image generation method is proposed in this paper.Firstly,defect-free images of polarizer is generated by histogram matching or encoding structured light,then the defect contour is constructed on the defect-free images of polarizer by using linear equation and elliptic curve equation,and four kinds of defect images are generated: bubble,crease,bump and foreign object.Finally,the polarizer image training set for training includes five types:bubble,crease,bump,foreign body and defect-free.Each type has 1800 generated images.There are 100 generated images and 100 real images for each type of image in the test set.The accuracy of the training set is 95.20%,and the accuracy of the test set is 94.60%.This shows that the trained five classification model can detect the defects of the real polarizer images.It also explains that the polarizer image generation method proposed is feasible.