外觀瑕疵檢測之研究

Research on surface defect detection

林侑萱(You-Hsuan Lin)

指導教授 ： 吳毅成

國立交通大學/資訊學院/資訊科學與工程研究所/碩士(2020年)

摘要

這篇論文提出一個檢測流程來偵測產品外觀的刮痕瑕疵，其中包含取像裝置研究、比對網路及去雜訊方法，使得檢測流程可以良好的檢測不易顯現的外觀刮痕瑕疵，並可以在沒有無瑕疵產品的情況下進行檢測。
我們觀察到在外觀瑕疵中，大部分刮痕瑕疵必須透過環境變化，如光線或相機角度的改變，才會清楚的顯現。因此，我們使用機器手臂夾取光源製造環境變化，並使用錄製影片的方式作為檢測資料。
結合深度學習技術，訓練模型需要大量的訓練資料，模型成果也會受限於訓練資料的品質及種類。我們提出了比對網路的概念，比對兩張照片的差異像素點來檢測的瑕疵。由於比對採用兩張照片比較差異，所訓練出的模型亦不受限於產品樣式，也使得我們的檢測流程對新產品有良好的適應性。
然而真實照片可能因為拍攝角度或物品材質，使得照片經過比對所得結果包含大量雜訊，而刮痕瑕疵的亦可能出現在大量的雜訊結果中，必須謹慎處理去雜訊的問題。我們改良傳統線段檢測，並搭配投票驗證的方法去雜訊，最終顯示我們提出的檢測流程，可以快速適應新產品，對不易顯現的刮痕瑕疵具有良好的檢測能力。

關鍵字

外觀檢測 ； 瑕疵檢測 ； 深度學習

並列摘要

In this thesis, we propose a surface defect detection process to detect the scratches, which includes a camera system, a network structure, and a denoising algorithm. Our surface defect detection process is able to detect scratches that are not highly visible to human eyes, and can be used even without having to demonstrate what a flawless product looks like.
We have observed that most of the scratches can only be clearly seen when there are changes in the environment, such as altering the light source or camera angle. Therefore, we use a moving robot arm carrying the light source to produce environmental changes, and record videos as the robot arm moves. The recorded video is then used to detect defects.
We propose contrastive network, a network structure that finds the difference between two photos, to detect defects. Because our algorithm compares two photos, it is product-agnostic, which makes our process easily adaptable to new objects.
However, real-world photos may contain a lot of noise. The scratches may be hidden inside a large number of noise pixels. It is necessary to perform noise removal carefully. We improved the traditional line detection method, and utilized voting verification for denoising. Finally, our detection process can quickly adapt to new products and detect scratches that are not easily visible.

並列關鍵字

surface defect detection ； defect detection ； deep learning

參考文獻

[1] Chin, R. T., & Harlow, C. A. (1982). Automated visual inspection: A survey. IEEE transactions on pattern analysis and machine intelligence, (6), 557-573.

[2] Chin, R. T. (1988). Automated visual inspection: 1981 to 1987. In Computer Vision, Graphics, and Image Processing, 41(3), 346-381.

[3] Dom, B. E., Brecher, V. H., Bonner, R., Batchelder, J. S., & Jaffe, R. S. (1988). The P300: A system for automatic patterned wafer inspection. Machine Vision and Applications, 1(4), 205-221.

[4] Bulnes, F. G., Usamentiaga, R., Garcia, D. F., & Molleda, J. (2016). An efficient method for defect detection during the manufacturing of web materials. Journal of Intelligent Manufacturing, 27(2), 431-445.

[5] Nakashima, K. (1994, May). Hybrid inspection system for LCD color filter panels. Conference Proceedings. 10th Anniversary. IMTC/94. Advanced Technologies in I & M. 1994 IEEE Instrumentation and Measurement Technolgy Conference (Cat. No. 94CH3424-9) (pp. 689-692). IEEE.