Surface Defect Detection And Classification Of Mobile Phone Screen Glass Based On Machine Vision

Mobile phone glass screen(MPSG)is one of the important components of the mobile phone,and its surface defects directly affect the quality of the mobile phone.Currently,the vibration of motion stage and illumination variation in capturing MPSG images cause image displacement and rotation,image gray variation and fuzzy boundary of defects,which decreases the defect detection accuracy.Meanwhile,high-dimensional defect features and imbalanced defects in MPSG surface will affect the accuracy of minority class defects in classification.These problems restrict the detection automation on the MPSG surface defects.In order to solve these problems,this thesis investigates the methods of defect identification,defect segmentation,defect feature extraction and imbalanced defect classification,and proposes a number of solutions,such as an image registration method based on contour similarity,an image subtraction and projection method for defect identification,an integrated FCM method based on spatial relation and membership correction,a multifractal spectrum defect feature extraction method,and a different contribution sampling method.With the algorithm implementation and system development,this thesis can realize the rapid detection and imbalanced classification to the MPSG surface defects.The main work of the thesis is summarized as follow.1)The thesis investigates thoroughly visual inspection and classification methods of MPSG surface defects.The current research situations and technological development are reviewed at home and abroad.The key technical problems are presented in MPSG defect detection and classification.2)The gray-based registration methods are time-consuming,and the Harris corner-based registration method easily suffers from the false corners.In order to solve these problems from the gray-based and the Harris corner-based registration methods,a new registration method is proposed based on similarity measure of MPSG surface contours.The new registration method quickly align the template image with the test image,which is required by the process of defect identification.In order to overcome the problem of illumination variation from image subtraction method in the defect detection,a method is proposed which combines the operation of subtraction with projection(CSP).The CSP method implements the projection algorithm based on the residual image after subtraction operation,thus eliminates the effects of illumination variation to the image gray.The experimental results show that the proposed CSP method achieves satisfying sensitivity and specifity.3)The fuzzy C-means clustering algorithm(FCM)is sensitive to noise in segmenting defects of the fuzzy boundary between defects and the background in the noisy images.In order to solve this problem,an integrated FCM(IFCM)method based on spatial relation and membership correction is proposed,which modifies the objective function of FCM algorithm using the spatial information of pixels,and also modifies the membership function of central pixels according to the membership degree of neighborhood pixels.The IFCM algorithm takes advantage of image gray information and spatial information,thus improves the anti-noise ability of FCM algorithm in the noisy defect image segmentation.4)In the defect feature extraction,high-dimension defect features easily cause the low classification accuracy and speed.Meanwhile,single fractal dimension can not characterize the complexity and inhomogeneity of defects well.Aiming to solve these problems caused by high-dimension defect features and single fractal dimension,multifractal spectrum features are extracted,which use multifractal spectrum difference?f and multifractal spectrum width Aa to describe the complexity and inheterogeneity of defects.However,the multifractal spectrum features are difficult to distinguish dirt defects from edge breakage defects.The location features of defects are extracted to solve this problem caused by the multifractal spectrum features.The experimental results show that the multifractal spectrum features and location features can distinguish the defects well.5)The existing sampling methods easily cause losing some important information,adding some trivial information and over-fitting problems in addressing the imbalanced data.In order to solve these problems,the thesis analyzes the different contribution from support vectors and non-support vectors to classifying the imbalanced data.Based on this analysis,a different contribution sampling(DCS)method is proposed.The DCS method uses the multiple random under-sampling technique(RUS)and the synthetic minority over-sampling technique(SMOTE)to re-sample the non-support vectors(NSVs)in the majority class and the support vectors(SVs)in the minority class respectively.Support vector machine ensembles are integrated with multiple support vector machine models and determine the labels of data through voting rules.Experimental results show that the proposed DCS method can improve the classification accuracy of minority defects without hurting the classification accuracy of majority defects.6)Based on the above-mentioned investigations into the related algorithms,the thesis sets up an experimental platform of MPSG image acquisition,designs and realizes the related algorithms and software development.The thesis also testifies the effectiveness and correctness of the algorithms proposed in MPSG defect detection and classification on some real MPSG images.