以影像幾何特徵建立液晶螢幕之MURA分類的決策樹模型

Mura是TFT-LCD顯示器上亮度不均勻造成的各種痕跡，可以使用自動光學檢測系統進行辨識，將Mura圖樣拍照存檔，而目前圖樣分類作業仍倚賴大量人工作業。人工作業常因主觀認知的不同，導致分類結果因人而異，更易因長期工作造成視覺疲勞進而形成誤判。因此，本研究針對Mura圖樣提出減少人力，提高作業速度與穩定性的自動分類技術。首先，將Mura影像二值化後計算Mura涵蓋區域的總面積、周長、長軸、短軸和分散度、平均亮度、亮度標準差、偏態係數、灰階變化率之以及內外面積比等幾何特徵，而後根據這些幾何特徵值利用決策樹分析將Mura圖樣進行分類。 188張Mura圖樣包含墨團、長條、分散、圓形以及亮度不均做實例分析，5種圖樣捕捉率和正確率之平均值分別為94.21%和93.57%，其中共有11張Mura圖樣分類錯誤。檢視這些這11張分類錯誤的圖樣，人工分類亦發生無法十分肯定的困境。故在驗證實驗時，我們排除這些圖樣和使用層別隨機抽樣評估決策樹的分類績效，經過1000次的抽樣模擬，整體捕捉率和正確率之平均值分別為96.52%和96.57%。是以，本研究使用幾何特徵對Mura圖樣進行自動化分類與人工辨識約有3%的差異，但因自動化加快分類速度，避免人工疲勞造成的分類錯誤和提升後續問題分析的能力，對於製程品質的提升將有顯著貢獻。

關鍵字

圖像辨識；二值化影像；決策樹；液晶面板

並列摘要

Mura is a visual defect due to non-uniformity on the surface of the backlight layer of TFT-LCD products. Now, mura can be identified and photted by automatic optical inspection system, however, it is still necessary for a lot of manpower to classify mura pictures into several clusters. Since the perceived differences between the inspectors, the classification results are easily inconsistent. Moreover, visual fatigue is likely to make the missing rate and false rate increase. Hence, the objective of this study is to develop the automatic mura classification algorithm by its several geometric characteristics. At first, we propose ten geometric characteristics of a mura body: size of mura, perimeter, maximum projective range, minimum projective range, splash, the average of grey level luminance, the standard deviation of grey level luminance, the skewness of grey level luminance, the proportion of within variation of two stratified muras, and the ratio of two stratified muras. Secondly, we use decision tree method to classify 188 real mura pictures into five clusters to investigate if the proposed geometric characteristic variables are adequate. There are five proposed geometric characteristics remained in the classification algorithm, the catch rate and accurate rate are 94.21% and 93.57%, respectively. Finally, we investigate the performance of the decision tree algorithm by the stratified cross validation method. After omitting 11 error classified pictures, the averages of catch rate and accurate rate in 1000 replications are 96.52% and 96.57%, respectively.