| Popis: |
The number of application areas of deep neural networks for image classification is continuously growing. A general desired attribute of these networks is to generalize well to test data that visually differs from the training data, but still shows the relevant features of the classes to be discriminated. Reasons for such a difference in data could be related to a change in background, illumination, or camera properties. The research area of Domain Adaptation (DA) deals with the transferability of classification models between such datasets, called domains, with the target to maximize the transferability. Typically, the differences and similarities of domains are described by the notion of general data distributions. This method, however, does not allow to identify and describe sufficiently the actual cause of a reduced performance on a new domain. To tackle this, in this thesis a novel description of domains, based on a theory of visual factors that describes the characteristics of domains will be introduced. As it will be shown, it can also be used to explain the targets and effects of existing DA approaches more understandable, which ultimately can be used to improve those even further. When it comes to the application of classification models in context of domains, several generalization cases can occur. In literature the most relevant ones are the cases where the application domain is the same domain as the training domain or the application domain is a completely new domain. The case that the application domain was one of multiple training domains is usually neglected, but will be investigated in this thesis as well, since it has high relevance for the usage of pre-trained classification models on own image data. As it will be shown further, the awareness about the domains for all three generalization cases is important for a well performing classification model in the application domain. The novel investigations in this context will be introduced under the term Effects of Domain Awareness. Different cases of domain awareness are investigated in combination with different domain constellations within the training and test data using the simple DA method of RGB mean normalization. The results on a road segmentation task show the importance to treat a domain during training and test always in the same way, since otherwise a significantly reduced performance can be observed. A typical assumption in current DA research is that each training domain includes samples for all classes that should be discriminated. However, thinking of distributed camera systems with a shared classification model, where each system potentially represents a domain, this assumption is too restricted. The more realistic assumption here is that not all classes are covered by samples from each domain during training of the classifier. The aforementioned scenario, which is overlooked in literature, will be extensively investigated under the term Domain Mixture scenario in this thesis. The experiments on MNIST and real-world object classification data show that, given the Domain Mixture scenario, the application of an approach from DA is essential, since otherwise the classification model is not capable to perform well on domain-class combinations that were not represented by supervised samples during training. A common DA approach to obtain a classification model that performs invariant of a domain well, is to remove all factors from the internal class feature representation that allow a discrimination of domains. This, however, can be harmful if at the same time task-informative factors are removed. To prevent this negative effect, the novel approach of Factor-Preserving DA (FP-DA) will be introduced which allows to preserve a selected factor during training with an adversarial DA approach. The experiments in this context will first show on real-world data that this negative effect exists and afterwards how factors worth preserving can be identified and subsequently be preserved through FP-DA in a multi-domain setting. The results show that FP-DA is capable to achieve the highest average and minimum performance in such a setting compared to the used baseline method. In summary, this thesis introduces a novel description of domains and based on that, investigates multiple highly relevant constellations for DA and additionally proposes a novel DA approach. |
