Wenyu Zhang

Advisors
Matthias Zürl (M.Sc.), Franz Koeferl (M.Sc.), Prof. Dr. Björn Eskofier, Christian Ott (Intego), Dr. Thomas Wagner (Intego)

Duration
11 / 2020 – 05 / 2021

Abstract
Anomaly detection and classification is an essential part in various medical diagnoses systems, e.g. tumor cells [1] or ECGs [2]. Therefore, applying deep learning techniques to detect and classify anomalies is a rapidly growing field in medical application settings. Understandably, medical data is subject to strict protection, which is why an analogous problem from industry is being tackled within the scope of this thesis. The statements made so far remain correct; for industrial applications, too, the detection and classification of anomalies using machine learning is becoming increasingly important. In manufacturing processes this enables an improvement of production pipelines and therefore increases the overall output of defect-free products.

The Intego GmbH – a medium-sized company located in Erlangen – develops and produces customer-specific inspection systems. They apply the latest measurement and imaging technologies to perform complex inspection tasks reliably [3]. Intego developed an optical inspection system for the Fraunhofer IISB to detect and classify defects on silicon carbide (SiC) wafers with aid of various sensors and traditional state of the art image processing techniques [4]. Their system uses ResNet [5] for the classification of defects on SiC-wafers, outperforming classical algorithms by a large margin. However, on images with a dierent manufacturing or inspection process, the performance of ResNet [5] trained on an original dataset decreases considerably. This is due to a change in the distribution of features and labels, i.e. defect classes, image contrast, etc.
In literature, this problem is called Domain Shift [6, 7, 8, 9]. The most commonly used method to solve this issue to fine-tune the networks parameters to fit the new data. Lee et al. [7] applied this technique on a similar industrial dataset called DAGM [10]. They outperformed their baseline model with frozen layers by 14.95%. Nevertheless, one disadvantage of fine-tuning is the requirement of labels on the new domain. Tzeng et al. [8] proposed an unsupervised domain adaptation approach using adversarial training on the problem of character classification and cross-modality adaptation. Kouw et al. [9] proposes the use of several known techniques like importance weighting or sub space mapping, depending on the kind of data shift.

This thesis applies domain adaptation techniques frequently used in literature to tackle the task of SiC-wafer defect classification. The goal is to achieve a classification performance in the target domain (shifted/new dataset) comparable to performance in the source domain (old/original dataset).

References:
[1] H. Ben Hsieh, Dena Marrinucci, Kelly Bethel, Douglas N. Curry, Mark Humphrey, Robert T. Krivacic, Joan Kroener, Lindsay Kroener, Andras Ladanyi, Nicole Lazarus, Peter Kuhn, Richard H. Bruce, Jorge Nieva, (2006),High speed detection of circulating tumor cells, Biosensors and Bioelectronics, Volume 21, Issue 10
[2] Haemwaan Sivaraks, Chotirat Ann Ratanamahatana (2015), Robust and Accurate Anomaly Detection in ECG Artifacts Using Time Series Motif Discovery, Computational and Mathematical Methods in Medicine
[3] Intego GmbH – https://www.intego.de/en/
[4] Doll, A (2018), Deep Learning Algorithmen im Leistungsvergleich zu konventionellen Klassifikationsverfahren an Beispielen der industriellen Sichtprufung, Master’s Thesis
[5] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2016-December, 770{778. https://doi.org/10.1109/CVPR.2016.90
[6] Y. Luo, L. Zheng, T. Guan, J. Yu and Y. Yang, “Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation,”2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019, pp. 2502-2511, doi: 10.1109/CVPR.2019.00261.
[7] Kim, S., Kim, W., Noh, Y. K., & Park, F. C. (2017). Transfer learning for automated optical inspection. Proceedings of the International Joint Conference on Neural Networks, 2017-May, 2517{2524. https://doi.org/10.1109/IJCNN.2017.7966162
[8] Tzeng, E., Homan, J., Saenko, K., & Darrell, T. (2017). Adversarial discriminative domain adaptation. Proceedings – 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, 2017-Janua, 2962{2971. https://doi.org/10.1109/CVPR.2017.316
[9] Kouw, W. M., & Loog, M. (2018). An introduction to domain adaptation and transfer learning. Retrieved from http://arxiv.org/abs/1812.11806
[10] DAGM Dataset of the DAGM 2007 Challenge, Retrieved from https://resources.mpiinf. mpg.de/conference/dagm/2007/prizes.html
[11] LeCun, Y., Cortes, C. (2010), ‘MNIST handwritten digit database’
[12] Baba Dash (2020). USPS Digit Dataset, MATLAB Central File Exchange
[13] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, Andrew Y. Ng Reading Digits in Natural Images with Unsupervised Feature Learning NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011.