Scalable Content-Based Classification and Retrieval Framework for Dynamic Commercial Image Databases

Abstract

Large-scale commercial image databases are getting increasingly common and popular, and nowadays several services over them are being offered via Internet. They are truly dynamic in nature where new image(s), categories and visual descriptors can be introduced in any time. In order to address this need, in this paper, we propose a scalable content- based classification and retrieval framework using a novel collective network of (evolutionary) binary classifier (CNBC) system to achieve high classification and content-based retrieval performances over commercial image repositories. The proposed CNBC framework is designed to cope up with incomplete training (ground truth) data and/or low-level features extracted in a dynamically varying image database and thus the system can be evolved incrementally to adapt the change immediately. Such a self-adaptation is achieved by basically adopting a “Divide and Conquer” type approach by allocating an individual network of binary classifiers (NBCs) to discriminate each image category and performing evolutionary search to find the optimal binary classifier (BC) in each NBC. Furthermore, by means of this approach, a large set of low-level visual features can be effectively used within CNBC, which in turn selects and combines them so as to achieve highest discrimination among each individual class. Experiments demonstrate a high classification accuracy and efficiency of the proposed framework over a large and dynamic commercial database using only low-level visual features.