利用極端梯度提升決策樹建構液晶面板自動光學檢測瑕疵分類模型

Abstract

TFT-LCD面板為許多電子產品的關鍵零組件之一，其品質亦決定了後續產品的價值。自動光學檢測為工業4.0中重要的一環，而TFT-LCD面板檢測也逐漸導入自動光學檢測技術以改善過去人力檢測的不足並提升其檢測準確率與效率。然而大部分文獻多針對面板上的刮痕、雲紋等瑕疵為主，較少針對壞點瑕疵進行，因此本研究以壞點瑕疵為目標，應用特殊濾鏡與局部二值模式(Local Binary Pattern)於影像前處理，並提出以簡單特徵向量做為極端梯度提升決策樹(XGBoost)演算法模型之輸入，建構出一套完整的TFT-LCD自動光學檢測瑕疵分類模型。利用台灣某公司之面板實際應向作為驗證資料，輸入本研究所建構的TFT-LCD瑕疵分類模型進行面板瑕疵之預測，得到97%的平均準確率。本研究亦與過去所使用GLCM(Gray Level Co-occurrence Matrix)特徵值的提升決策樹分類器進行效能比較，得到本研究所提出之分類模型在面板壞點瑕疵檢測上有著更快的檢測速度與效能的結論。

Thin film transistor liquid crystal display (TFT-LCD) panel is a key component in many electronic products. Its quality determines the value of follow-up products. Automatic optical inspection (AOI) plays an important role in Industry 4.0 and it has been introduced to screen out bad TFT-LCD panels during quality inspection. However most of the exist literature related to applying AOI in TFT-LCD panel inspection focus on defects like scratches, Mura, particles, etc. Few of them studied defective pixels on TFT-LCD panel. Therefore, this study aims to apply special filter and local binary pattern (LBP) in the image processing, and to extract a set of features as the input for applying the eXtreme Gradient Boosting Decision Tree (XGBoost) as the classifier to sort panels with defective pixels out from good ones. Real images of TFT-LCD panels provided by a computer manufacturer in Taiwan panel are used to verify the effectiveness and efficiency of the proposed approach. The results of numerical experiment showed that the proposed approach outperformed the ones used in previous studies in terms of average classification accuracy as well as time efficiency.