利用具注意力機制之生成對抗網路於工業影像之瑕疵單一分類

One-Class Defect Classification of Industrial Images Using Generative Adversarial Network with Attention Mechanism

羅倆凡(Liang-Fan Luo)

指導教授 ： 胡竹生

國立交通大學/電機學院/電機與控制工程系所/碩士(2019年)

摘要

在工業製造過程中，瑕疵檢測是不可或缺的一環。大部分檢測瑕疵的方式須先定義瑕疵，找到符合該瑕疵的特徵，最後在設定適當的閾值來做判斷。而現今的檢測流程中也存在各式懸而未決的問題。瑕疵的多樣性、工業取像環境的變化等造成的問題，使得模型的強健性遭受挑戰；類別不平衡及瑕疵樣本的稀缺性會造成標註及訓練定義上的困難；而瑕疵的定義及瑕疵的標註則會耗費大量的人力資源及時間。

受到這些問題的激勵，本篇論文發展出一套新的檢測機制：利用生成對抗網路(generative adversarial networks, GANs)生成工業影像，並在生成影像的過程中，獲取無瑕疵的工業影像在隱空間中的分布，以此為分類的標準，判斷新輸入的工業影像是否有瑕疵。其中導入了注意力機制於我們的模型中，使得經由生成器生成出來的影像，能與真實的影像更相近，在隱空間上的分布也會越逼近真實影像；再者，輸入到模型的影像亦經過了灰階值上的增量與標準化等前處理，期望分類結果能不受光影的不一致所影響。本篇論文中所提方法在開源的工業數據庫DAGM的第八類及第十類上皆有良好的分類結果，於實際工業生產線上的偏光片影像上測試也有不錯的分類結果。

關鍵字

自動光學檢測 ； 生成對抗模型 ； 單一類別分類模型

並列摘要

Inspecting defects is an indispensable part in the industrial manufacturing process. Most defect-inspection methods follow the steps: defining defects, finding features that conform to the detected defects and then classifying by a proper threshold. However, there are still many unsolved problems in the contemporary inspecting process. For example, defect variety and different image-capturing environment seriously affect the robustness of inspecting models; unbalance and sparsity of defective samples cause difficulty in labeling and training; defining and labeling defects would consume many time and human resources.

In this thesis, we develop a new inspection mechanism to enhance the robustness and reduce the human effort in labeling. We obtain the distribution of non-defective industrial images in latent space during the process of generating images by generative adversarial networks (GANs). Then the non-defective distribution is used as the standard of classification to inspect the new testing data. To minimize the intensity variation of the acquired images, we also introduce an attention mechanism in the model to ensure the generating images and the real image are as similar as possible. Consequently, the distributions in the latent space would be similar as well. Before applying our method, the images are pre-processed for augmentation and standardization in their gray level to overcome the inconsistence of brightness. The proposed method significantly improve the classification results not only in class 8 and class10 of DAGM open source but also in the polarizer images in real industrial manufacturing lines.

並列關鍵字

AOI ； GANs ； One-Class classification

參考文獻

第一章

[1.1] R. Ren, T. Hung and K. C. Tan, “A Generic Deep-Learning-Based Approach for Automated Surface Inspection,” in IEEE Trans. on Cybernetics, vol. 48, no. 3, pp. 929-940, 2018.

[1.2] S. Kim, W. Kim, Y. K. Noh, and F. C. Park, “Transfer Learning for Automated Optical Inspection,” 2017 International Joint Conference on Neural Networks (IJCNN), pp. 2517-2524, 2017.

[1.3] M. Wieler and T. Hahn, “Weakly Supervised Learning for Industrial Optical Inspection,” available accessed at https://goo.gl/nEF3yQ. Retrieved at 30/04/2018.

[1.4] J. Ritcher, D. Streitferdt and E. Rozova, “On the development of intelligent optical inspections,” 2017 IEEE 7th Annual Computing and Communication Workshop and Conference (CCWC), pp. 1-6, 2017.