A new integrated SVM classifiers for relevance feedback content-based image retrieval using EM parameter estimation

Abstract

In content-based image retrieval (CBIR), the support vector machine (SVM) based relevance feedback is studied extensively to narrow the gap between low-level image feature and high-level semantic concept. Despite the success, for conventional SVM relevance feedback, the retrieval performance is actually worse when the number of labeled positive feedback samples is small. To overcome this limitation, a SVM classifier combination for relevance feedback content-based image retrieval using expectation–maximization (EM) parameter estimation is proposed. Firstly, we introduce the asymmetric bagging SVM to improve the stability of SVM classifiers and the balance in the training. Then, the random subspace SVM is used to overcome the overfitting problem. Finally, we combine the asymmetric bagging SVM and the random subspace SVM using EM parameter estimation to form an integrated SVM as a hypothesized solution to the overall image retrieval problem, which can further improve the relevance feedback performance. Experiments on large databases show that the proposed algorithms are significantly more effective than the state-of-the-art approaches.