UNIFIED LEARNING PARADIGM FOR IMAGE RETRIEVAL

作者：Lin, Zheng-Kui; Wu, Jun^*; Xiao, Zhi-Bo; Duan, Jing; Lu, Ming-Yu

来源：International Journal of Innovative Computing Information and Control, 2011, 7(8): 4977-4987.

• Content-based image retrieval
• Relevance feedback
• Short-term learning
• Long-term learning
• Unified learning

Dealing with Relevance feedback (RF) using statistical learning has been a key technique to improve the content-based image retrieval (CBIR) performance. However, there is still a big room to further RF performance since the popular RF methods ignore the cooperation among various learning mechanisms. In this paper, we propose a unified learning paradigm (ULP) that integrates the merits of ensemble learning, semi-supervised learning, active learning and long-term learning into a uniform framework. Concretely, unlabeled examples are exploited to facilitate ensemble learning by helping augment the diversity among the base classifiers, and then, a strong ensemble is used to identify the most informative examples for active learning. In particular, the semantic clues are inferred in the long-term learning setting, which serves as the prior knowledge to validate the effectiveness of the unlabeled examples used by ULP. Finally, a bias-weighting strategy is developed to guide the ensemble of classifiers to pay more attention to the positive examples than the negative ones. An empirical study shows that using multiple learning strategies simultaneously in CBIR is beneficial, and that the proposed scheme is significantly more effective than some existing approaches.

• 出版日期2011-8
• 单位大连海事大学