Relevance feedback using semi-supervised learning algorithm for image retrieval

Relevance feedback (RF) based on support vector machine (SVM) has been widely used in content-based image retrieval (CBIR) to bridge the semantic gap between low-level visual features and high-level human perception. However, the conventional SVM based RF uses only the labeled images for learning, which gives rise to the small sample problem, i.e., when the training data is insufficient, the performance of SVM may drop dramatically. In this paper, we alleviate the small sample problem in SVM based RF by adopting semi-supervised active learning algorithm that builds better models with a large amount of unlabeled data and the labeled data. Active learning is used to alleviate the manual effort for labeling by selecting only the informative data. In addition, a semi-supervised approach has been developed, which employs Bayesian classifier to label the data with a certain degree of uncertainty in its class information. Using these automatically labeled samples, fuzzy support vector machine (FSVM) is trained, which takes into account the fuzzy nature of some training samples. We compare our method with standard active SVM based RF on a database of 10,000 images, the experimental results show that our method has a better performance and effectiveness on the CBIR task.