為了確保產品品質，表面瑕疵檢測在許多不同的行業中皆扮演著至關重要的角色。傳統的手動檢查方法通常耗時且主觀，可能導致不准確和更高的生產成本。使用深度學習技術於表面瑕疵檢測在近年來取得了顯著進展。此外，深度學習包含各種技術，而其中，圖像分類的深度學習技術與我們的研究領域密切相關，也是本研究的重點。
在工業表面缺陷檢測領域中，研究者們一直致力於創建一種基於深度學習的智能缺陷檢測系統，該系統希望能實現接近零缺陷率，同時保持輕量、高效和低成本的解決方案。然而，這些目標經常相互衝突，開發一個能夠同時實現所有目標的模型是不現實的，通常必須做出一些取捨。舉例而言，如果精度是首要考慮因素，則可能需要取得大量且標記過的瑕疵圖像，並進行監督式學習。而如果優先考慮輕量級和低成本，則通常會使用簡單的小型模型，例如Auto-Encoder，並使用大量的無瑕疵圖像進行非監督式學習，以最大程度地減少標記成本。
如前所述，設計一個能夠同時實現所有目標的單一模型非常困難。然而，現今的研究通常集中於使用單一模型完成這些任務，並很少涉及多模型架構。本文提出了一種表面缺陷檢測和分類系統，該系統主要由三個模型組成，分別是以CNN為基礎的表面缺陷檢測領域中當前最先進的模型[1]，以及基於VAEGAN的零樣本學習分類器，還有由我們實驗室開發的Variational Auto-Encoder模型[2]組成。
我們研發了一套表面瑕疵檢測和分類系統，該系統有效地整合了上述三種模型。它已在金屬表面缺陷數據集上得到驗證，並得到了優秀的結果。此系統不僅實現了符合工業等級標準的瑕疵檢出率和低誤報率，同時還保持了輕量級、低延遲和低標記成本等特性。除了實現上述目標外，此系統還可以在出現未見過的異常時立即識別和發出異常通知，這在絕大部分使用監督式學習的異常檢測模型中是不可能實現的。

Surface defect detection is crucial in maintaining product quality across various industries. Traditional manual inspection methods are often time-consuming and subjective, which can result in inaccuracies and higher production costs. With the use of deep learning techniques, significant advancements have been made in automating the process of surface defect detection in recent years. Moreover, deep learning includes a variety of techniques, and image recognition-based deep learning is especially relevant to our field of study, which is the main focus of this research paper.
In the industrial surface defect detection field, researchers have always aimed to create a deep learning-based intelligent defect detection system that achieves near-zero defect rates while maintaining a lightweight, efficient, and cost-effective solution. However, these objectives often conflict with each other, and it is unrealistic to develop a model that can achieve all of them simultaneously. Some trade-offs must be made. If accuracy is the top priority, a large amount of defective data labeled for supervised learning is usually required. If lightweight and low cost is prioritized, a simple small model such as Auto-Encoder is usually used, along with a large number of flawless images for unsupervised learning to minimize the cost of labeling.
As mentioned before, it is very difficult to design a single model that can achieve all of them simultaneously. However, present-day studies frequently center on accomplishing those tasks using a single model and rarely address the multi-model architecture. This paper presents a Surface Defect Detection and Classification System that builds on the current state-of-the-art model [1] in the field of surface defect detection, along with the zero-shot learning (ZSL) classifier based on VAEGAN and the Variational Auto-Encoder developed by our laboratory [2].
We have developed a Surface Defect Detection and Classification System that effectively integrates the aforementioned three models. It has been validated on a dataset of metal surface defects, yielding excellent results. This system not only achieves defect detection rates that comply with industrial standards and low false positive rates but also maintains characteristics such as lightweight, low latency, and low annotation cost. In addition to achieving the above goals, this system can also instantly identify and issue anomaly notifications when there are unseen anomalies, which is generally impossible to do with supervised learning anomaly detection models.

1. Božič, J., D. Tabernik, and D. Skočaj. End-to-end training of a two-stage neural network for defect detection. in 2020 25th International Conference on Pattern Recognition (ICPR). 2021. IEEE.
2. Pan, C.-C., Evolution of Class Embedding for Unseen Class in Zero-Shot Learning. 2021.
3. Cohn, R. and E. Holm, Unsupervised machine learning via transfer learning and k-means clustering to classify materials image data. Integrating Materials and Manufacturing Innovation, 2021. 10(2): p. 231-244.
4. Vahdat, A. and J. Kautz, NVAE: A deep hierarchical variational autoencoder. Advances in neural information processing systems, 2020. 33: p. 19667-19679.
5. Chandola, V., A. Banerjee, and V. Kumar, Anomaly detection: A survey. ACM computing surveys (CSUR), 2009. 41(3): p. 1-58.
6. Božič, J., D. Tabernik, and D. Skočaj, Mixed supervision for surface-defect detection: From weakly to fully supervised learning. Computers in Industry, 2021. 129: p. 103459.
7. Tabernik, D., et al., Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 2020. 31(3): p. 759-776.
8. Li, J., et al., Real-time detection of steel strip surface defects based on improved yolo detection network. IFAC-PapersOnLine, 2018. 51(21): p. 76-81.
9. Guo, Z., et al., Msft-yolo: Improved yolov5 based on transformer for detecting defects of steel surface. Sensors, 2022. 22(9): p. 3467.
10. Goodfellow, I., et al., Generative adversarial nets. Advances in neural information processing systems, 2014. 27.
11. Xian, Y., et al. Feature generating networks for zero-shot learning. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.
12. Arjovsky, M., S. Chintala, and L. Bottou. Wasserstein generative adversarial networks. in International conference on machine learning. 2017. PMLR.
13. Gao, Y., L. Gao, and X. Li, A generative adversarial network based deep learning method for low-quality defect image reconstruction and recognition. IEEE Transactions on Industrial Informatics, 2020. 17(5): p. 3231-3240.
14. Narayan, S., et al. Latent embedding feedback and discriminative features for zero-shot classification. in European Conference on Computer Vision. 2020. Springer.
15. Zhao, Z., et al. A surface defect detection method based on positive samples. in PRICAI 2018: Trends in Artificial Intelligence: 15th Pacific Rim International Conference on Artificial Intelligence, Nanjing, China, August 28–31, 2018, Proceedings, Part II 15. 2018. Springer.
16. Wu, H. and Q. Lv, Hot-rolled steel strip surface inspection based on transfer learning model. Journal of Sensors, 2021. 2021: p. 1-8.
17. Weng, L., Lil'Log, in From Autoencoder to Beta-VAE. 2018.
18. Liu, J., et al., Semi-supervised anomaly detection with dual prototypes autoencoder for industrial surface inspection. Optics and Lasers in Engineering, 2021. 136: p. 106324.
19. Kingma, D.P. and M. Welling, Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
20. He, K., et al. Deep residual learning for image recognition. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
21. Xian, Y., et al., Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 2018. 41(9): p. 2251-2265.
22. Catherine Wah, Steve Branson, Peter Welinder, Pietro Per-ona, and Serge Belongie. The caltech-ucsd birds-200-2011dataset. 2011.
23. Weimer, D., B. Scholz-Reiter, and M. Shpitalni, Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection. CIRP annals, 2016. 65(1): p. 417-420.
24. Simonyan, K. and A. Zisserman, Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
25. Ding, X., et al. Repvgg: Making vgg-style convnets great again. in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.