(599j) Evaluation of Classification Algorithms for Automatic Inspection of TFT-LCD Glass Substrates

This paper presents an industrial case study of automatic inspection methodology to manufacturing of TFT-LCD glass substrates. We perform a three-way comparison of classification accuracy involving classification & regression tree (CART), multi-layer perceptron (MLP) and support vector machine (SVM). Decision making was performed in four stages: feature extraction by computation of wavelet co-occurrence signature, handling imbalanced data problem through the synthetic minority over-sampling technique (SMOTE), dimensional reduction with principal component analysis (PCA), and classification by using three different classification methods (CART, MLP and SVM optimized via simulated annealing algorithm (SA)).

The more SMOTE we used, the higher the accuracy we obtained by each three models. Optimized SVM and MLP models classified TFT-LCD glass substrates more accurately than the CART model. The SVM model was better in classification than the MLP model. The accuracy for the classification using CART and optimized MLP were 62.9% and 87.9%, respectively while a higher accuracy of 89.5% was observed for the optimized SVM model. This study shows possibility of automation of glass inspection in glass substrates manufacturing industry.