在目前傳統的螺絲的自動光學檢測(Automated Optical Inspection, AOI)，大多是針對瑕疵特性開發客製化演算來抓出螺絲影像異常的區塊，但依不同表面處理製程產製的螺絲，其影像的差異非常大，因此為了讓演算法的通用性變高，通常會提供許多的參數來讓使用者依每批螺絲的情況來微調演算法，如此使用者的技術水平需求會變高，除了需要了解各個參數的意義，當一些新的異常型態出現時，這種針對性的演算法則效能會降低。因此本論文希望透過深度學習方法來簡化使用者的負擔以及加強對未知異常型態的偵測能力，並利用螺絲正常品的取得在螺絲產業相對容易的特性，讓使用者只需用正常品圖片就可訓練深度學習模型以偵測螺絲缺陷品。
本研究以四種生成網路模型來比較分析，分別為卷積自動編碼器(Convolutional Auto-Encoder, CAE)、變分自動編碼器(Variational Auto-Encoder, VAE)、使用生成對抗網路之變分自動編碼器(Variational Auto-Encoder with Generative Adversarial NetworkVAEGAN)、及使用生成對抗網路之快速非監督式異常偵測(Fast Unsupervised Anomaly Detection with Generative Adversarial Networks,f-AnoGAN)四種生成網路模型，方法為透過螺絲正常品圖片來訓練各模型，讓模型只學習螺絲正常品圖片的特徵，當輸入螺絲異常品圖片時，模型能生成相似的正常品圖片，藉由比較輸入圖與模型產出圖之間的差異並設定門檻值來達到螺絲異常品偵測。之後選出以對異常品數據敏感度最高之模型來實驗其在不同型號螺絲異常品偵測的表現。最後四種中以 CAE 模型表現較佳，該 CAE 模型可在短時間內訓練出較好的特徵擷取器，並將螺絲異常品圖片轉為較接近正常品螺絲之圖片，因此易於偵測，而其在其他型號的螺絲上重新訓練後，也有不錯的螺絲異常品偵測表現。

Most traditional Automated Optical Inspection (AOI) screw defects detection systems are customized for detecting flaws of screws made with specific surface processing methods. To deal with different methods, which delivers screws whose optical images varies widely, designers usually setup a lot of system arguments for users to adapt for different batches of screw produced with different methods. Such strategy increases users’technique load because they must understand well the meaning and influence of each argument and, when a new method is introduced to the production line, it might reduce the defects detection performance of a system. This research aims to reduce users’ load and construct a screw image defects detection neural network system that can be easily retrained in a short time when new method is introduced. This study evaluates four models: Convolutional AutoEncoder (CAE), Variation Auto-Encoder (VAE), Variation Auto-Encoder with Generative Adversarial Network (VAEGAN), and Fast Unsupervised Anomaly Detection with Generative Adversarial Network (f-AnoGAN). These models are trained by flawless screw images which can be collected easily when new methods are used for producing screws. These models learn the features of normal screw images and, when reading in a screw image, with or without defects, they try to generate a normal screw image. The comparison between the original input image and the generated image yields a differential value that, when larger than a threshold value, indicates defects are in the original input screw image. The evaluations finds that the CAE model is most sensitive to screw defects and this study uses for further experiments using other types of screws to ensure that the CAE, after trained, also has the capability in detecting defects of screws produced with new methods. The experiments confirm that CAE model, after retrained for different types of screw images, can reach 100% precision in detecting defects of screws with a recall rate up to 99.6%.

Bank, D., Koenigstein, N., & Giryes, R. (2021). Autoencoders, Preprint arXiv:2003.05991v2.
Baur, C., Denner, S., Wiestler, B., Navab, N., & Albarqouni, S. (2021). Autoencoders for unsupervised anomaly segmentation in brain MR images: A comparative study. Journal of Medical Image Analysis, 69, 16.
Chen, D., Yue, L., Chang, X., Xu, M., & Jia, T. (2021). NM-GAN: Noise-modulated generative adversarial network for video anomaly detection. Journal of Pattern Recognition, 116.
Chen, S., Yu, J., & Wang, S. (2020). One-dimensional convolutional auto-encoder-based feature learning for fault diagnosis of multivariate processes. Journal of Process Control, 87, 54-67.
Goubeaud, M., Jousen, P., Gmyrek, N., Ghorban, F., Schelkes, L., & Kummert, A. (2021). Using Variational Autoencoder to augment Sparse Time series Datasets. The 7th International Conference on Optimization and Applications, ICOA 2021, May 19, 2021 - May 20, 2021, Wolfenbuttel, Germany.
Han, C., Rundo, L., Murao, K., Noguchi, T., Shimahara, Y., Milacski, Z. a., . . . Satoh, S. (2021). MADGAN: unsupervised medical anomaly detection GAN using multiple adjacent brain MRI slice reconstruction. Journal of BMC Bioinformatics, 22.
Kim, M. S., Yun, J. P., Lee, S., & Park, P. (2019). Unsupervised Anomaly detection of LM Guide Using Variational Autoencoder. The 11th International Symposium on Advanced Topics in Electrical Engineering, ATEE 2019, March 28, 2019 - March 30, 2019, Bucharest, Romania.
Kingma, D. P., & Welling, M. (2014). Auto-encoding variational bayes. The 2nd International Conference on Learning Representations, ICLR 2014, April 14, 2014 - April 16, 2014, Banff, AB, Canada.
Larsen, A. B. L., Sonderby, S. K., Larochelle, H., & Winther, O. (2016). Autoencoding beyond pixels using a learned similarity metric. The 33rd International Conference on Machine Learning, ICML 2016, June 19, 2016 - June 24, 2016, New York City, NY, United states.
Liu, X., Gherbi, A., Wei, Z., Li, W., & Cheriet, M. (2021). Multispectral Image Reconstruction from Color Images Using Enhanced Variational Autoencoder and Generative Adversarial Network. Journal of IEEE Access, 9, 1666-1679.
Nguyen, M.-H., Nguyen, D.-Q., Nguyen, D.-Q., Pham, C.-N., Bui, D., & Han, H.-D. (2021). Deep Convolutional Variational Autoencoder for Anomalous Sound Detection. The 8th IEEE International Conference on Communications and Electronics, ICCE 2020, January 13, 2021 - January 15, 2021, Phu Quoc, Viet nam.
Nho, Y.-H., Ryu, S., & Kwon, D.-S. (2021). UI-GAN: Generative Adversarial Network-Based Anomaly Detection Using User Initial Information for Wearable Devices. IEEE Sensors Journal, 21(8), 9949-9958.
Principi, E., Rossetti, D., Squartini, S., & Piazza, F. (2019). Unsupervised Electric Motor Fault Detection by Using Deep Autoencoders. Ieee-Caa Journal of Automatica Sinica, 6(2), 441-451.
Schlegl, T., Seebock, P., Waldstein, S. M., Langs, G., & Schmidt-Erfurth, U. (2019). f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks. Journal of Medical Image Analysis, 54, 30-44.
Wang, T., Liu, J., Jin, C., Li, J., & Ma, S. (2020). An intelligent music generation based on Variational Autoencoder. The 2020 International Conference on Culture-Oriented Science and Technology, ICCST 2020, October 30, 2020 - October 31, 2020, Beijing, China.
Wang, X., & Liu, H. (2020). Data supplement for a soft sensor using a new generative model based on a variational autoencoder and Wasserstein GAN. Journal of Process Control, 85, 91-99.
Xiao, N., Qiang, Y., Zhao, Z., Zhao, J., & Lian, J. (2020). Tumour growth prediction of followup lung cancer via conditional recurrent variational autoencoder. Journal of IET Image Processing, 14(15), 3975-3981.
Yang, D., Karimi, H. R., & Sun, K. (2021). Residual wide-kernel deep convolutional autoencoder for intelligent rotating machinery fault diagnosis with limited samples. Journal of Neural Networks, 141, 133-144.
Yokkampon, U., Chumkamon, S., Mowshowitz, A., & Hayashi, E. (2020). Anomaly Detection using Variational Autoencoder with Spectrum Analysis for Time Series Data. The Joint 9th International Conference on Informatics, Electronics and Vision and 4th International Conference on Imaging, Vision and Pattern Recognition, ICIEV and icIVPR 2020, August 26, 2020 - August 29, 2020, Kitakyushu, Japan.
Zhang, R., Bian, S., Liu, Y., & Li, H. (2019). Widespread bathymetric outliers detection and elimination based on conditional variational autoencoder generative adversarial network. Journal of Cehui Xuebao/Acta Geodaetica et Cartographica Sinica, 48(9), 1182-1189.
Zhang, Y. (2018). A Better Autoencoder for Image: Convolutional Autoencoder, Australian National University Paper Canberra ACT 2601, Australia.
高逸軒. (2018). 扣件螺紋瑕疵檢測-SVM 與 CNN 之分類效能比較. (碩士論文), 南臺科技大學, 台南市.
陳彥霖. (2015). 機器視覺於螺絲頭部記號篩選之分析. (碩士論文), 國立高雄海洋科技大學, 高雄市.
陳思翰. (2010). 以機器視覺為基礎之內螺紋自動檢測系統. (博士論文), 國立交通大學,新竹市.
黃將威. (2013). 一種創新式外螺紋表面呈現的機械視覺技術. (碩士論文), 明志科技大學, 新北市.
翟靖華. (2010). 應用於輪緣螺帽分檢及扣件表面缺陷檢測之電腦視覺系統. (碩士論文), 國立成功大學, 台南市.