Group-based relevance feedback with support vector machine ensembles

Support vector machines (SVMs) have become one of the most promising techniques for relevance feedback in content-based image retrieval (CBIR). Typical SVM-based relevance feedback techniques simply apply the strict binary classifications: positive (relevant) class and negative (irrelevant) class. However, in a real-world relevance feedback task, it is more reasonable and practical to assume the data come from multiple positive classes and one negative class. In order to formulate an effective relevance feedback algorithm, we propose a novel group-based relevance feedback scheme constructed with the SVM ensembles technique. Experiments are conducted to evaluate the performance of our proposed scheme and the traditional SVM-based relevance feedback technique in CBIR. The experimental results show that our proposed scheme is more effective than the regular method.