Active sample-selecting and manifold learning-based relevance feedback method for synthetic aperture radar image retrieval

Abstract

Content-based image retrieval (CBIR) provides an effective way to address the increasing need for intelligent data access to synthetic aperture radar (SAR) image repositories. In CBIR, a critical component is relevance feedback (RF), which is used to bridge the ‘semantic gap’. This study proposes a new RF method with active sample-selecting and manifold learning for CBIR of SAR images. In this method, the authors adopt a modified maximum margin projection with a new neighbourhood estimation criterion to discover both the geometrical structure and discriminant structure of the underlying data manifold. In order to achieve a satisfactory performance with limited feedback samples, the authors also propose an active sample selection strategy with which the diversity of feedback samples can be increased while the redundancy is decreased. The authors test our method on a TerraSAR-X image database and compare it with four other state-of-the-art RF methods. The superiority and validity of the authors’ method is proved by the retrieval results and the computing cost is acceptable for image retrieval applications.