- Titel
- Applications of Invertible Neural Networks: From Outlier Detection to Solving Inverse Problems
- Zitierfähige Url:
- https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-953223
- Erstveröffentlichung
- 2025
- Datum der Einreichung
- 27.08.2024
- Datum der Verteidigung
- 27.11.2024
- Abstract (EN)
- Deep learning algorithms often face limitations in terms of their reliability in domains such as autonomous driving and medical imaging. Although these models have achieved notable success in solving inverse problems, their typical unidirectional mapping can still limit their performance. Invertible Neural Networks (INNs) can address these shortcomings, thanks to their unique ability to maintain invertibility between input and output paradigms while estimating complex data distributions. This thesis investigates the application of INNs to address three distinct research problems. The first research problem focuses on developing robust methodologies for outlier detection within the computer vision tasks of image classification and object detection. To tackle this problem, I developed novel methods, each of which learned the inlier data distribution using INNs. For incorporating outlier awareness in the object detection pipeline in particular, I generated synthetic outlier data from the reverse direction of the learned INN model for contrastive training. The method was rigorously validated against existing state-of-the-art approaches, and it demonstrated superior performance in identifying outliers while even reducing false positives. The second research problem involves solving an inverse problem in current tomography by utilizing a simulation setup that mimics electrolysis for bubble detection, which is crucial for efficient hydrogen production. An INN-based model is developed to reconstruct higher-resolution conductivity maps from lower-resolution magnetic flux density measurements. The results show that compared to traditional linear models, INNs can more robustly reconstruct these maps against changes in the noise level of magnetic flux density between training and inference. The third research problem addresses the non-decomposability of conventional image fusion models. An INN based model has been developed that not only fuses source medical images but also decomposes the fused image back into the original source images, a task that was previously unexplored. All in all, these three research problems may seem distinct, but the proposed INN-based methods provide a unified approach to address them, advancing each domain and paving the way for future research and broader applications of INNs in other research domains.
- Freie Schlagwörter (DE)
- Maschinelles Lernen, Computer Vision, invertierbare neuronale Netze, medizinische Bildgebung, autonomes Fahren, Stromtomographie
- Freie Schlagwörter (EN)
- Machine Learning, Computer Vision, Invertible Neural Networks, Medical Imaging, Autonomous Driving, Current Tomography
- Klassifikation (DDC)
- 006
- Klassifikation (RVK)
- ST 301
- ST 330
- ST 640
- ST 620
- GutachterIn
- Prof. Dr. Stefan Gumhold
- Prof. Dr. Carsten Rother
- BetreuerIn Hochschule / Universität
- Prof. Dr. Stefan Gumhold
- Den akademischen Grad verleihende / prüfende Institution
- Technische Universität Dresden, Dresden
- Version / Begutachtungsstatus
- publizierte Version / Verlagsversion
- URN Qucosa
- urn:nbn:de:bsz:14-qucosa2-953223
- Veröffentlichungsdatum Qucosa
- 05.02.2025
- Dokumenttyp
- Dissertation
- Sprache des Dokumentes
- Englisch
- Lizenz / Rechtehinweis
CC BY 4.0