Research On Defects Imaging,Extraction,Recognition And Classification For TFT-LCD Panels

The size of thin film transistor-liquid crystal display(TFT-LCD)panels is becoming larger and larger with the development of panel industry,the traditional human visual inspection is gradually replaced by the automatic optical inspection which is regarded as one of the most important technology in improving the efficiency and quality in the production.This article mainly studies the defect extraction,the extraction and the selection of defect features and the multi-classification.In the defect extraction,the experiment result of the saliency detection model based on the spectral residual showed that the model was sensitive to the defect sizes,so this model was improved and applied to the detection of the high-energy frequency component in the spectrum which was obtained from the panel image by two-dimensional discrete Fourier transform(DFT).That component was set to zero by the frequency domain filter which was generated by the improved model,and then the spatial domain image was reconstructed by the two-dimensional inverse discrete Fourier transform(IDFT)in order to remove the background texture and retain the defect.The parameters of the model were determined by the gray level co-occurrence matrix(GLCM)of the image,the improved model was not sensitive to the defect sizes,the defect grayscale and the background texture under the fixed parameters.In the defect feature extraction,some algorithms were optimized with the defect images in this article.The two-pass algorithm in the binary image connected components labeling was optimized by reducing the sort operation.Compared with the original algorithm,the speed was accelerated to 179.3%.In the convex hull part,the Andrew's algorithm and the Melkman's algorithm were tested and the edge points of defects were reduced by a novel method which was proposed in this article.The speed of the proposed method which reduced the edge points in the horizontal direction was 41.6% faster than that of the original method,the speed of the Andrew's algorithm was 11.2% faster than that of the Melkman's algorithm.The improved rotating calipers algorithm was applied to get the minimum-area encasing rectangles(MER)of the defects and it was 13.3% faster than the function in the opencv,the morphological characteristic and the moment of the defects were also calculated.In the classification part,the defects were classified to 4 classes according to their shape features.In order to reduce the computational complexity in training support vector machine(SVM),the features of defects were filtered by the standard deviation,the Pearson correlation and the distance correlation.The libsvm toolbox was applied to create the one-versus-one(OVO)single kernel SVM and the hierarchical single kernel SVM which based on the characteristic of defects.The optimal parameters in the single kernel SVM classifiers are obtained by 5-fold cross validation.The SMO-MKL toolbox was applied to create the hierarchical multiple kernel SVM which based on the characteristic of defects.The experiment result shows that the hierarchical multiple kernel SVM which based on the characteristic of defects has the best performance.