發光二極體(Light emitting diode, LED)具壽命長、耗電量低、體積小等優良特性,近年大量被使用在各種用途上,造就了光電相關產品的發展,本研究主要將針對表面具複雜結構性紋路之LED透鏡進行可視瑕疵的偵測,目前許多LED透鏡之生產過程的品質檢測,仍然仰賴人工使用目視的方式進行瑕疵檢測,在持續長時間工作下,易造成眼睛疲勞,導致品質檢測上產生較多的誤判發生,造成損失。有鑑於此,本研究將試圖建構一套LED透鏡表面可視瑕疵之自動化檢測系統,主要提出使用小波包轉換 (Wavelet Packet Transform, WPT) 擷取紋路與瑕疵之小波包特徵,搭配使用多變量分析中的部份最小平方 (Partial Least Square, PLS) 法,將小波包特徵向量進行轉換與維度縮減,得到小波包特徵中重要的資訊,可以有效地將背景中紋路的影響降至最低,並將瑕疵檢測出來。此方法相較於以往刪除整個小波子區塊的方式,更能保留表面紋路與瑕疵資訊,而實驗結果顯示在瑕疵檢出率可達93.5 %,而瑕疵誤判率僅0.102%,執行每張影像檢測時間僅頇113毫秒(ms),非常適用於實際生產作業中進行瑕疵的檢測。
This research proposes a wavelet packet transformation (WPT) based partial least square (PLS) approach to detect visual defects of optical components with structural textures. Three steps are developed to finish the process of defect detection. Firstly, a spatial domain image is converted to WPT domain and the wavelet features of the sub-band images are extracted. Secondly, the proposed PLS method is applied to multivariate transform and data reduction with wavelets features to obtain latent images. There is as much information in the latent components as those in the original features. Thirdly, the latent images are fitted by a regression model to produce a predicted image and then subtract with the original image to get the residual image where the visual defects have been separated. Experimental results show that performance of the wavelet based PLS approach in the defect detection rate is 93.5% better than that of PCA model.