Deep attribute networks

Abstract

Representing an image as compact and discriminative features is one of the major challenges in many of the real-world image classification tasks in computer vision. One possible approach is to represent an image with high-level descriptors which are semantic and compact representation of the image. The term, high-level descriptors can be considered as human description which humans tend to use when classifying objects. In real-world applications, we often face ambiguous classification problems, involving tremendous variations between intra class objects. However, humans can handily classify the objects by understanding the scenes, which is a starting point of the intelligence. Humans can understand the scene because they use high-level description that carries semantic meaning of visual information. To be specific, humans have the ability to describe an object with attributes, characterizing its visual aspects into lists of words. Composed sets of attributes gain discriminative ability, replacing low-level descriptors in real-world classification. Based on this paradigm, a two-step framework can be considered in a classification task. First, obtain an attribute descriptor (binary state or real-valued score) which is semantic and compact representations of given images. Then classification is performed by a final classifier, which takes the obtained attribute descriptor as its input. Since the attribute descriptor is abstraction of low-level descriptors, it often presents better performance. In this thesis, two novel attribute classifiers based on deep learning are proposed. The efficacy of the proposed models is evaluated on unconstrained face verification and real-world object recognition tasks using the LFW and the a-PASCAL datasets. The potential of deep learning for attribute-based classification was successfully demonstrated by showing comparable results with existing state-of-the-art results. The efficiency of deep learning is that once properly trained ...