Feature Engineering in Fine-Grained Image Classification

Abstract

In most machine learning pipelines, feature engineering is an important module. The use of features has become the bottleneck for many learning related tasks, especially fo unstructured data like images and texts. This thesis researched on the use of features from several aspects, including feature design, feature selection, and feature learning, on a family of tasks called fine-grained image classification. Fine-grained classification refers to tasks in which the class differences are very subtle to observe, for example, to recognize sub-ordinate level classes like certain animal species, or to identify similar 3D shapes in medical images. In particular, this thesis covers several different projects. The use of feature selection algorithm is first explored in analysing similar 3D shapes for a medical problem called craniosynostosis. Different sparse logistic regression models are investigated, and a new sparse logistic regression model called clustering lasso is proposed specifically for this problem. Next, on a specific fine-grained recognition problem -- fast food recognition, an image representation called pairwise feature distribution (PFD) is proposed, which is focused on capturing the spatial information inside food images, using geometric pairwise features. The use of feature learning approaches is then explored on the general fine-grained object recognition problem, and a template model is proposed to improve the state-of-the-art object recognition framework by learning of mid-level feature representations for fine-grained tasks. The effectiveness of these algorithms proposed in this thesis is shown by comparison with the state-of-the-art algorithms on several publicly available benchmark datasets.