An Active-Learning approach to the query by example retrieval in remote sensing images

Abstract

In this paper, we propose an Active Learning approach to query by example retrieval, using a retraining procedure that improves the understanding of the machine with respect to the human perception. The proposed method is based on Support Vector Machine (SVM) classifiers and requires a small number of training samples. The classifier is retrained several times in order to determine the optimal separating hyper-plane between the class of the query and the rest of the analysed image. The closest feature points to the SVM-learned hyper-plane are the points being able to produce the most relevant modification of the position of this hyper-plane. These points, that are both negative and positive examples, are then used to retrain the SVM classifier. In addition, the proposed approach shows the importance of normalization in a classification problem with heterogeneous objects. Several experiments were conducted on GeoEye-1 multispectral images, whilst the retrieval was performed for different patch-level descriptors, which furthermore increases the complexity of the semantic content of the query object.