| Thin flm transistor liquid crystal display (TFT-LCD) has been widely used dueto its color performance, energy saving and small size. Defect inspection and classif-cation are very important for TFT-LCD manufacture, which would improve the yieldand quality and reduce costs. Inspection of TFT-LCD defect has been investigated alot. But there is not much rearch on defect classifcation.The topic discussed in this article is a5-class defect classifcation problem. Eachsamplefromdatasetcontains3images: originalimage, defectregionimageandcircuitzoneimage. Wearegoingtodevelopaclassiferbasedoncomputervisionandmachinelearning. Theinputis defectsampleand theoutputispredictedclasslabel. Developingan efective solution from scratch is a big challenge, since previous studies are verylimited.In order to solve the problem of defect classifcation, we explored a variety ofmethods. At frst, we experimented on some simple image features, including shapefeature, gray histogram and LBP (Local Binary Pattern). Then we applied the methodof visual bag of words, which is originally used in natural image classifcation. Wegave a detailed experiment result on color descriptor, image patch descriptor and SIFTdescriptor with diferent parameters and varieties. We also investigated the results ofNa ve Bayes, linear kernel SVM and χ2kernel SVM classifers and combinationalclassifers with diferent features. The experiment results show that the χ2kernel SVMclassifer with dense keypoints and combinational feature of color and SIFT achievesa good performence.At last we presented the implement of the algorithm using C++on Windows. Wedesigned a software framework based on plain text data, small programs and Make system. It has a lot of advantages: loose coupling in interface and data structure, easeto modify and debug, parallel operation. It is not only designed for algorithm of clas-sifcation based on visual bag of words, but also can be used in any machine learningsoftware on single computer. |
PDF Full Text Request