The use of unlabeled data in image retrieval with relevance feedback

This paper describes a content-based image retrieval (CBIR) system which makes use of both labeled images, annotated by the user, and unlabeled images available in the database. The system initially retrieves images objectively closest to the query image. The user then subjectively labels retrieved images as relevant or irrelevant. Although such relevance feedback from the user is an effective way of bridging the semantic gap between objective and subjective similarity, it is also very time consuming, requiring huge human effort. Often, the number of labeled images is very small. In an inductive approach the labeled set of images is used for training a CBIR system while the large set of unlabeled images remains unused. In this paper we exploit the transductive support vector machine (SVM) algorithm as a way of taking advantage of unlabeled data in CBIR. Our findings are compared to the results of an inductive SVM. We draw some conclusions as to when the use of unlabeled data might be helpful. The considered systems are tested over images from the Corel 1K dataset.