Classification Methods Of Defects On TFT-LCD Panels

This article relies on the National Special Major Instrument" Flat Panel Display Automatic Optical Inspection Instrument Development and Application", to study the detection of the defects, feature selection and feature extraction, and to study the classification methods of point, line and Mura defect on TFT-LCD panel scanning images.According to the periodic property of TFT-LCD panel,1D-FFT is used to remove the periodic background, while it brings into the boundary effect, a one-dimensional splicing principle is used to solve the problem. For the uneven illumination problem, full size wavelet transform is used to provide a sufficient condition for the threshold segmentation. Finally, this article eliminates the false defect, the defects are detected accurately.The hoteling transform is used to mark the minimum rectangle of defect, the shape features are acquired. Then the three statistical moment features are computed. So the four kinds of features are selected as classification features. To reduce redundant data, we use PCA to improve the processing speed, meanwhile the validity are ensured. After feature extraction, the amount of computation is reduced and the computation speed has been greatly improved.Experienced humans are used to finish TFT-LCD defect classification in most of TFT-LCD companies, So a method based on rules according to personnel discrimination rules is proposed in this article. The support vector machine(SVM) method has excellent performance on classification problem. To increase the validity of classification method, this article introduces the SVM. This paper uses traditional kernel SVM classification to classify defects based on libSVM toolbox developed by National Taiwan University professor Chi-Jen Lin. In order to make the classifier more universal, more stable, a 2-kernel SVM classifier and 4-kernel SVM classifier based on multi-kernel are cunstructed in this article. Finally, the several classifiers are compared in accuracy. The experiment result shows that the 2-kernel SVM has the best performance.