Exploiting Unlabeled Data in Content-Based Image Retrieval

Abstract

Abstract. In this paper, the Ssair (Semi-Supervised Active Image Retrieval) approach, which attempts to exploit unlabeled data to improve the performance of content-based image retrieval (Cbir), is proposed. This approach combines the merits of semi-supervised learning and active learning. In detail, in each round of relevance feedback, two simple learners are trained from the labeled data, i.e. images from user query and user feedback. Each learner then classifies the unlabeled images in the database and passes the most relevant/irrelevant images to the other learner. After re-training with the additional labeled data, the learners classify the images in the database again and then their classifications are merged. Images judged to be relevant with high confidence are returned as the retrieval result, while these judged with low confidence are put into the pool which is used in the next round of relevance feedback. Experiments show that semi-supervised learning and active learning mechanisms are both beneficial to Cbir.

Semisupervised Regression with Cotraining-Style Algorithms

“…Another co-training style algorithm which uses more than two learners has been presented by Zhou and Goldman [47]. Some variants of co-training [48], [49] which combine semi-supervised learning with active learning and do not require different views, have been applied to content-based image retrieval, where images provided by the user in the query and relevance feedbacks are regarded as labeled examples while the images existing in the image database are regarded as unlabeled examples.…”

“…This implies that the conditional independence [8] or even the weak dependence [2] between the two views is not needed, at least, for iterative co-training which is actually the working routine taken by many co-training style algorithms [20], [47], [48], [49], [51]. In fact, the assumption of two sufficient views is too strong that Zhou et al [53] have shown that when this assumption can be met, semi-supervised learning given only one labeled example is feasible.…”

“…Such a setting can also bring another advantage, that is, since it is usually difficult to decide which distance metric and which k value is better for the concerned task, the functions of these regressors may be somewhat complementary if they are combined. Note that the use of different distance metrics has been shown helpful in some variants of co-training [48], [49].…”

“…While in semisupervised learning, since there is no oracle that can be relied, the unlabeled example on which the learner is with the most confidence is usually selected to label. These two learning processes have been combined in some previous work [48], [49].…”

Locating Regions of Interest in CBIR with Multi-instance Learning Techniques

“…Here the main difficulty lies in the gap between the high-level image semantics and the low-level image features, due to the rich content but subjective semantics of an image. Although much endeavor has been devoted to bridging this gap [4] [11], it remains an unsolved problem at present. Nevertheless, many good CBIR systems have already been developed.…”