Semi-supervised learning for image classification

Abstract

Object class recognition is an active topic in computer vision still presenting many challenges. In most approaches, this task is addressed by supervised learning algorithms that need a large quantity of labels to perform well. This leads either to small datasets (< 10,000 images) that capture only a subset of the real-world class distribution (but with a controlled and verified labeling procedure), or to large datasets that are more representative but also add more label noise. Therefore, semi-supervised learning is a promising direction. It requires only few labels while simultaneously making use of the vast amount of images available today. We address object class recognition with semi-supervised learning. These algorithms depend on the underlying structure given by the data, the image description, and the similarity measure, and the quality of the labels. This insight leads to the main research questions of this thesis: Is the structure given by labeled and unlabeled data more important than the algorithm itself? Can we improve this neighborhood structure by a better similarity metric or with more representative unlabeled data? Is there a connection between the quality of labels and the overall performance and how can we get more representative labels? We answer all these questions, i.e., we provide an extensive evaluation, we propose several graph improvements, and we introduce a novel active learning framework to get more representative labels.