A novel active learning method in relevance feedback for content based remote sensing image retrieval
Download as DOCX, PDF0 likes171 views
This document presents a novel active learning method to improve content-based remote sensing image retrieval with reduced user labeling effort. The method selects the most informative images for relevance feedback by jointly evaluating images based on uncertainty, diversity, and density in the image archive. First, the most uncertain images are selected using a margin sampling strategy. Then, diverse and high-density region images are chosen from the uncertain images using a novel clustering-based strategy. Experimental results demonstrate the effectiveness of the active learning method at driving relevance feedback to retrieve remote sensing images from large archives.
A novel active learning method in relevance feedback for content based remote sensing image retrieval
- 1. A NOVEL ACTIVE LEARNING METHOD IN RELEVANCE FEEDBACK FOR CONTENT-BASED REMOTE SENSING IMAGE RETRIEVAL ABSTRACT Conventional relevance feedback (RF) schemes improvethe performance of content- based image retrieval (CBIR)requiring the user to annotate a large number of images. To reducethe labeling effort of the user, this paper presents a novel activelearning (AL) method to drive RF for retrieving remote sensingimages from large archives in the framework of the supportvector machine classifier. The proposed AL method is specificallydesigned for CBIR and defines an effective and as small as possibleset of relevant and irrelevant images with regard to a generalquery image by jointly evaluating three criteria: 1) uncertainty;2) diversity; and 3) density of images in the archive. The uncertaintyand diversity criteria aim at selecting the most informativeimages in the archive, whereas the density criterion goal isto choose the images that are representative of the underlyingdistribution of data in the archive. The proposed AL method assesses jointly the three criteria based on two successive steps.In the first step, the most uncertain (i.e., ambiguous) images areselected from the archive on the basis of the margin samplingstrategy. In the second step, the images that are both diverse (i.e.,distant) to each other and associated to the high-density regionsof the image feature space in the archive are chosen from the mostuncertain images. This step is achieved by a novel clustering-basedstrategy. The proposed AL method for driving the RF contributesto mitigate problems of unbalanced and biased set of relevant andirrelevant images. Experimental results show the effectiveness ofthe proposed AL method.