Deep Learning Techniques for Anomaly and Defect Detection Applications
Event details
| Date | 21.06.2019 |
| Hour | 09:00 › 11:00 |
| Speaker | David Honzatko |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Martin Jaggi
Thesis advisor: Prof. Pascal Fua
Co-examiner: Dr. Martin Rajman
Abstract
Application of deep learning based computer vision algorithms in industry is often problematic due to their need for a large annotated datasets.
Particularly demanding task is detection of defects and anomalies in production.
Not only is the annotation process expensive, but also it might be infeasible to obtain representative data for the annotation.
In a typical scenario, there are hundreds of genuine samples and tens of defective ones. For each of these there are multiple observations.
In this work, we investigate defect and anomaly detection architectures that could benefit from multiple observations, capture the defining features of the genuine samples, and make use of the limited amount of annotations available.
Background papers
Anomaly Detection: A Survey, by Chandola, Varun, Arindam Banerjee, and Vipin Kumar. ACM Computing Surveys (CSUR) 41.3 (2009): 1-58.
Pixel recurrent neural networks, by Van Den Oord, Aäron, Nal Kalchbrenner, and Koray Kavukcuoglu. Proceedings of the 33rd International Conference on on Machine Learning-Volume 48. JMLR. org, 2016.
Deformable Convolutional Networks, by Dai, Jifeng, et al. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017.
Exam president: Prof. Martin Jaggi
Thesis advisor: Prof. Pascal Fua
Co-examiner: Dr. Martin Rajman
Abstract
Application of deep learning based computer vision algorithms in industry is often problematic due to their need for a large annotated datasets.
Particularly demanding task is detection of defects and anomalies in production.
Not only is the annotation process expensive, but also it might be infeasible to obtain representative data for the annotation.
In a typical scenario, there are hundreds of genuine samples and tens of defective ones. For each of these there are multiple observations.
In this work, we investigate defect and anomaly detection architectures that could benefit from multiple observations, capture the defining features of the genuine samples, and make use of the limited amount of annotations available.
Background papers
Anomaly Detection: A Survey, by Chandola, Varun, Arindam Banerjee, and Vipin Kumar. ACM Computing Surveys (CSUR) 41.3 (2009): 1-58.
Pixel recurrent neural networks, by Van Den Oord, Aäron, Nal Kalchbrenner, and Koray Kavukcuoglu. Proceedings of the 33rd International Conference on on Machine Learning-Volume 48. JMLR. org, 2016.
Deformable Convolutional Networks, by Dai, Jifeng, et al. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017.
Practical information
- General public
- Free