산업용 이미지의 구조적 차이를 식별하는 이상 탐지

국문 초록 (Abstract)

제조 설비 기술의 발달로 산업현장에서 불량 데이터 확보가 점차 어려워 데이터 불균형을 야기한다. 머신비전 분야에서는 데이터 불균형 문제를 해결할 수 있는 Autoencoder, GAN 방식의 Unsupervised 이상 탐지 기술이 각광받고 있다. 하지만 Autoencoder, GAN 방식은 학습이 불안정하며, 일반적으로 단순 픽셀 차이를 계산하는 L1, L2 손실 함수를 사용하여 구조적 차이를 판단하기 어렵다. 따라서 동일 클래스에서 비정상 데이터를 탐지하는 Novelty Detection 이상 탐지 분야에 적합하지 않다. 본 연구에서는 복원력을 향상시킬 수 있는 U-Net 기반 복원 네트워크를 사용하고, 동일 클래스에서 Local Region의 구조적 차이를 확인할 수 있는 SSIM, MS-SSIM 손실 함수를 사용한 방법을 제시한다. 해당 방법은 실제 산업현장에서 발생하는 물체의 결함을 판단하는데 사용하는 제조 분야에 특화된 MVTec 데이터를 사용해서 실험을 진행하였고 Autoencoder, GAN 방식 네트워크에서 L1, SSIM, MS-SSIM 손실 함수를 사용하여 성능 비교 실험하였다.

다국어 초록 (Multilingual Abstract)

Due to the development of manufacturing technology, it is increasingly difficult to gather data flaws in the manufacturing process, leading to imbalanced data. Therefore, Autoencoder and GAN-based unsupervised anomaly detections that solve the problem of imbalanced data are in the spotlight. However, the Autoencoder and GAN-based models are not only unstable during training, but also typically use L1 and L2 loss functions for a per-pixel error, making it harder to distinguish structural differences between input and reconstruction. Consequently, detecting anomalies in the same class is not suitable in novelty anomaly detection fields. In this work, we present a way to take advantage of SSIM, MS-SSIM loss functions on a U-Net network to identify structural differences between local regions in the same class and improve reconstruction stability while training. Utilizing MVTec dataset, which is specialized for manufacturing processes to detect defects, we compared the performances of L1, SSIM, MS-SSIM on Autoencoder, GAN-based networks.

목차 (Table of Contents)

• 요약
• Abstract
• Ⅰ. 서론
• Ⅱ. 관련 연구
• Ⅲ. 본론
• Ⅳ. 실험
• Ⅴ. 결론
• REFERENCES

참고문헌 (Reference)

1 Radford, Alec, "Unsupervised representation learning with deep convolutional generative adversarial networks"

2 Schlegl, Thomas, "Unsupervised anomaly detection with generative adversarial networks to guide marker discovery" Springer 2017

3 Ronneberger, Olaf, "U-net: Convolutional networks for biomedical image segmentation" Springer 2015

4 Akçay, Samet, "Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly detection" IEEE 2019

5 Hinton, Geoffrey E, "Reducing the dimensionality of data with neural networks" 313 (313): 504-507, 2006

6 Bergmann, Paul, "MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection" 2019

7 Bergmann, Paul, "MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection" 2019

8 Zhao, Hang, "Loss functions for image restoration with neural networks" 3 (3): 47-57, 2016

9 Liu, Peidong, "Loss Function Discovery for Object Detection via Convergence-Simulation Driven Search"

10 Isola, Phillip, "Image-to-image translation with conditional adversarial networks" 2017

11 Wang, Zhou, "Image quality assessment:from error visibility to structural similarity" 13 (13): 600-612, 2004

12 Akcay, Samet, "Ganomaly: Semi-supervised anomaly detection via adversarial training" Springer 2018

13 Chen, Po-Yu, "Distributed real-time anomaly detection in networked industrial sensing systems" 62 (62): 3832-3842, 2014

14 Ruff, Lukas, "Deep semi-supervised anomaly detection"

15 Fernando, Tharindu, "Deep Learning for Medical Anomaly Detection--A Survey"

16 Kingma, Diederik P., "Auto-encoding variational bayes"

17 Anandakrishnan, Archana, "Anomaly detection in finance: editors’ introduction" PMLR 2017