An ontological bagging approach for image classification of crowdsourced data

Abstract

In this paper, we study how to use semantic relationships for image classification in order to improve the classification accuracy. We achieve the goal by imitating the human visual system which classifies categories from coarse to fine grains based on different visual features. We propose an ontological bagging algorithm where most discriminative weak attributes are automatically learned for different semantic levels by multiple instance learning and the bagging idea is applied to reduce the error propagations of hierarchical classifiers. We also leverage ontological knowledge to augment crowdsourcing annotations (e.g., a hatchback is also a vehicle) in order to train hierarchical classifiers. Our method is tested on a vehicle dataset from the popular crowdsourcing dataset ImageNet. Experimental results show that our method not only achieves state-of-the-art results but also identifies semantically meaningful visual features.