異常偵測主要的目標是偵測出偏離正常分佈的樣本。我們這個方法主要應用在工業影像的異常偵測，用來偵測工業產品是否帶有瑕疵，並檢測出異常部分。因為異常圖片的不可預測性，蒐集與標記異常圖片也過於費工費時，所以現階段多數模型都將異常偵測以非監督式學習 (unsupervised learning) 的方式來訓練。近年來，許多方法在訓練中加入和真實異常情況相近的人工合成異常圖片，幫助模型在訓練過程中，針對分辨正常資料與異常資料的目的優化，進而協助模型更準確地學習正常資料的分佈。我們提出一個基於標準化流 (normalizing flow)的自監督學習方法，用來最大化估計正常資料的概似函數，並最小化人工異常資料的概似。我們從現有正常圖片切下數個隨機區塊，再將區塊融合進另一張正常圖片，藉此產生出與實際異常情況相似的人工異常圖片。另外，我們針對異常樣本在損失函數上加入額外條件限制，讓模型可以專注在優化正常樣本，也不容易被極端異常情況干擾訓練。我們堆疊卷積網路與自注意力機制，並加入殘差連接，優化了標準化流模型中的耦合 (coupling layer)。我們在公開資料集MVTec-AD, BTAD與DAGM的實驗結果顯示，我們提出的方法不論在用完整樣本學習或是小樣本學習，在異常偵測與分割的任務中，皆能達到當前最好的準確度。

Image anomaly detection aims to detect out-of-distribution instances. Most existing methods treat anomaly detection as an unsupervised task because anomalous training data and labels are usually scarce or unavailable. Recently, image synthesis has been used to generate anomalous samples which deviate from normal sample distribution for model training. By using the synthesized anomalous training samples, we present a novel self-supervised normalizing flow-based density estimation model, which is trained by maximizing the likelihood of normal images and minimizing the likelihood of synthetic anomalous images. By adding constraints to abnormal samples in our loss function, our model training is focused on normal samples rather than synthetic samples. Moreover, we improve the transformation subnet of the affine coupling layers in our flow-based model by dynamic stacking convolution and self-attention blocks. We evaluate our method on MVTec-AD, BTAD and DAGM datasets and achieve state-of-art performance on both the anomaly detection and localization tasks.

1 Introduction 1
1.1 Problem Statement 1
1.2 Motivation 2
1.3 Contributions 3
1.4 Thesis Organization 3

2 Related Work 5
2.1 Anomaly Detection Methods 5
2.1.1 Unsupervised Anomaly Detection 5
2.1.2 Self-supervised Anomaly Detection 6
2.1.3 Supervised Anomaly Detection 7
2.2 Normalizing Flow 7

3 Proposed Method 9
3.1 Self-supervised Learning 9
3.2 Normalizing Flows Architecture 10
3.2.1 Residual Connected Subnet 12
3.3 Learning Objective 13
3.4 Scoring Function 15

4 Experiments 17
4.1 Experimental Settings 17
4.1.1 Datasets and Evaluation Metric 17
4.1.2 Implementation Details 18
4.2 Experimental Comparisons 18
4.2.1 Results on MVTec-AD 19
4.2.2 Results on BTAD Dataset 20
4.2.3 Results on DAGM 20
4.3 Few-shot Anomaly Detection and Segmentation 25
4.4 Qualitative Results 29
4.5 Complexity Analysis 30

5 Ablation Study 32
5.1 Impact of Out-of-distribution Loss 32
5.2 Impact of Different Subnet and Residual Components 33

6 Conclusions 36
References 37

[1] Akcay, S., Atapour-Abarghouei, A., and Breckon, T. P. Ganomaly: Semi- supervised anomaly detection via adversarial training. In Asian conference on computer vision (2018), Springer, pp. 622–637.
[2] Akçay, S., Atapour-Abarghouei, A., and Breckon, T. P. Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly detection. In 2019 International Joint Conference on Neural Networks (IJCNN) (2019), IEEE, pp. 1–8.
[3] Ardizzone, L., Bungert, T., Draxler, F., Köthe, U., Kruse, J., Schmier, R., and Sorrenson, P. Framework for Easily Invertible Architectures (FrEIA), 2018- 2022.
[4] Bergmann, P., Fauser, M., Sattlegger, D., and Steger, C. Mvtec ad–a compre- hensive real-world dataset for unsupervised anomaly detection. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition (2019), pp. 9592–9600.
[5] Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., and Steger, C. Improv- ing unsupervised defect segmentation by applying structural similarity to au- toencoders. In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2019), SCITEPRESS - Science and Technology Publications.
[6] Božič, J., Tabernik, D., and Skočaj, D. End-to-end training of a two-stage neural network for defect detection. In 2020 25th International Conference on Pattern Recognition (ICPR) (2021), IEEE, pp. 5619–5626.
[7] Defard, T., Setkov, A., Loesch, A., and Audigier, R. Padim: a patch distribu- tion modeling framework for anomaly detection and localization. In Interna- tional Conference on Pattern Recognition (2021), Springer, pp. 475–489.
[8] Ding, C., Pang, G., and Shen, C. Catching both gray and black swans: Open- set supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 7388–7398.
[9] Dinh, L., Krueger, D., and Bengio, Y. Nice: Non-linear independent compo- nents estimation. arXiv preprint arXiv:1410.8516 (2014).
[10] Dinh, L., Sohl-Dickstein, J., and Bengio, S. Density estimation using real nvp. arXiv preprint arXiv:1605.08803 (2016).

[11] Dohi, K., Endo, T., Purohit, H., Tanabe, R., and Kawaguchi, Y. Flow-based self-supervised density estimation for anomalous sound detection. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2021), IEEE, pp. 336–340.
[12] Gudovskiy,D.,Ishizaka,S.,andKozuka,K.CFLOW-AD:Real-timeunsuper- vised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision (2022), pp. 98–107.
[13] He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016).
[14] Ho, J., Chen, X., Srinivas, A., Duan, Y., and Abbeel, P. Flow++: Improving flow-based generative models with variational dequantization and architecture design. In International Conference on Machine Learning (2019), PMLR, pp. 2722–2730.
[15] Jewell, J. T., Khazaie, V. R., and Mohsenzadeh, Y. One-class learned encoder- decoder network with adversarial context masking for novelty detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (January 2022), pp. 3591–3601.
[16] Kingma, D. P., and Dhariwal, P. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems 31 (2018).
[17] Kirichenko, P., Izmailov, P., and Wilson, A. G. Why normalizing flows fail to detect out-of-distribution data. Advances in neural information processing systems 33 (2020), 20578–20589.
[18] Li, C.-L., Sohn, K., Yoon, J., and Pfister, T. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (2021), pp. 9664–9674.
[19] Liang, Y., Zhang, J., Zhao, S., Wu, R., Liu, Y., and Pan, S. Omni-frequency channel-selection representations for unsupervised anomaly detection. arXiv preprint arXiv:2203.00259 (2022).
[20] Lin, Z., Ye, H., Zhan, B., and Huang, X. An efficient network for surface defect detection. Applied Sciences 10, 17 (2020), 6085.
[21] Matthias Wieler, Tobias Hahn, F. A. H. Weakly supervised learning for industrial optical inspection. https://hci.iwr.uni-heidelberg.de/content/ weakly-supervised-learning-industrial-optical-inspection, 2007.
[22] Mishra, P., Verk, R., Fornasier, D., Piciarelli, C., and Foresti, G. L. VT-ADL: A vision transformer network for image anomaly detection and localization. In 30th IEEE/IES International Symposium on Industrial Electronics (ISIE) (June 2021).
[23] Pang, G., Ding, C., Shen, C., and Hengel, A. v. d. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462 (2021).
[24] Pirnay, J., and Chai, K. Inpainting transformer for anomaly detection. In International Conference on Image Analysis and Processing (2022), Springer, pp. 394–406.
[25] Pourreza, M., Mohammadi, B., Khaki, M., Bouindour, S., Snoussi, H., and Sabokrou, M. G2d: generate to detect anomaly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2021), pp. 2003–2012.
[26] Rudolph, M., Wandt, B., and Rosenhahn, B. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision (2021), pp. 1907–1916.
[27] Rudolph, M., Wehrbein, T., Rosenhahn, B., and Wandt, B. Fully convolu- tional cross-scale-flows for image-based defect detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022), pp. 1088–1097.
[28] Runinho. Implementation of cutpaste. https://github.com/Runinho/ pytorch-cutpaste, 2022.
[29] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. Imagenet large scale visual recognition challenge. International journal of computer vision 115, 3 (2015), 211–252.
[30] Salehi,M.,Eftekhar,A.,Sadjadi,N.,Rohban,M.H.,andRabiee,H.R.Puzzle- ae: Novelty detection in images through solving puzzles. arXiv preprint arXiv:2008.12959 (2020).
[31] Schlegl, T., Seeböck, P., Waldstein, S. M., Langs, G., and Schmidt-Erfurth, U. f-anogan: Fast unsupervised anomaly detection with generative adversarial networks. Medical image analysis 54 (2019), 30–44.
[32] Schlegl, T., Seeböck, P., Waldstein, S. M., Schmidt-Erfurth, U., and Langs, G. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging (2017), Springer, pp. 146–157.
[33] Schlüter, H. M., Tan, J., Hou, B., and Kainz, B. Self-supervised out-of- distribution detection and localization with natural synthetic anomalies (nsa). In European conference on computer vision (2022).
[34] Sharma, R., Mashkaria, S., and Awate, S. P. A semi-supervised generalized vae framework for abnormality detection using one-class classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022), pp. 595–603.
[35] Song, J., Kong, K., Park, Y.-I., Kim, S.-G., and Kang, S.-J. Anoseg: Anomaly segmentation network using self-supervised learning. arXiv preprint arXiv:2110.03396 (2021).
[36] Tan, J., Hou, B., Batten, J., Qiu, H., and Kainz, B. Detecting outliers with foreign patch interpolation. arXiv preprint arXiv:2011.04197 (2020).
[37] Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., and Jégou, H. Going deeper with image transformers. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (2021), pp. 32–42.
[38] Tsai, C.-C., Wu, T.-H., and Lai, S.-H. Multi-scale patch-based representation learning for image anomaly detection and segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022), pp. 3992–4000.
[39] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems 30 (2017).
[40] Yi, J., and Yoon, S. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian Conference on Computer Vision (ACCV) (November 2020).
[41] Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., and Wu, L. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677 (2021).
[42] Zavrtanik, V., Kristan, M., and Skočaj, D. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2021), pp. 8330– 8339.
[43] Zavrtanik, V., Kristan, M., and Skočaj, D. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition 112 (2021), 107706.