Guided transfer learning approach for domain adaptation in image classification and object detection

Abstract

Deep learning has achieved remarkable success in various computer vision tasks, such as image classification and object detection. However, these tasks typically require a massive amount of labeled data to train deep Convolutional Neural Network (CNN) models and learn complex patterns, which is often a challenging and time-consuming process. Transfer learning has emerged as a prevalent technique that leverages pre-trained models to alleviate the data requirement by fine-tuning the models on the target dataset. However, the effectiveness of transfer learning depends on the similarity between the source and target domains and the availability of labeled data in the target domain, which is not always the case in many real-world scenarios. Furthermore, domain-specific tasks generally perform well when the feature distributions of the domains are similar. The feature distributions can differ from one domain to another due to various factors such as intra-class variations, camera sensor variations, background variations, illumination changes, and geographical differences. Consequently, utilizing a trained source model directly in the target domain may not generalize effectively, even when the label space of both domains is the same. As a result, the models are more likely to misclassify unknown classes as known classes and vice versa, which leads to the domain shift problem. Domain adaptation has emerged as a transfer learning approach to address the domain shift problem by reducing the gap between the domains. However, a challenge arises while using transfer learning, which is how deep to fine-tune the CNNs while training the target model to maximize the transferability. Additionally, reducing the domain gap with the unlabeled target domain poses another significant challenge. newlineThe aim of this thesis is to design and implement an efficient domain adaptation network to learn transferable feature representations and reduce the domain shift problem in a unified network with labeled source data and unlabeled target data.