Learning mid-level features from object hierarchy for image classification

Abstract

One of the most active research areas in computer vision is image classification.Although there have been many research efforts in this area, it remains a difficult problem, especially when the number of categories is large. Most of the previous work in image classification uses low-level image features. We believe low-level features ignore a lot of the semantic structures of the image classes. In this thesis, we go beyond simple low-level features and propose new approaches for constructing midlevel visual features for image classification. We represent an image using the outputs of a collection of binary classifiers. These binary classifiers are trained to differentiate pairs of object classes in an object hierarchy. Our feature representations implicitly capture the hiererarchical structure in object classes. We show that our proposed approach outperforms other baseline methods in image classification.ii

Object Classification Using a Semantic Hierarchy

“…Our work is most closely related to [12] and [13]. Cao et al [12] proposed a framework for learning mid-level features called "learning by focus".…”

“…The reason for using one-vs-one (instead of one-vs-rest) classifiers is that it is easier to distinguish different concept pairs. Albaradei et al [13] extended [12] by learning binary classifiers for concept pairs at different levels of the object hierarchy. The main difference between our work and [13] is how to construct image representations from those binary classifiers.…”

“…Albaradei et al [13] extended [12] by learning binary classifiers for concept pairs at different levels of the object hierarchy. The main difference between our work and [13] is how to construct image representations from those binary classifiers. The method in [13] chooses to concatenate the scores of all binary classifiers as the image representation.…”

“…The main difference between our work and [13] is how to construct image representations from those binary classifiers. The method in [13] chooses to concatenate the scores of all binary classifiers as the image representation. In contrast, our image representation will exploit the hierarchical structure of the object categories.…”

“…For example, [12,13] apply a large collection of binary classifers on the low-level image features. The scores of these binary classifiers are used as the mid-level features.…”