An efficient CBIR system for high resolution remote sensing images

Abstract

High resolution satellite images (HRSI), contain a great range of objects and spatial patterns appearing with a large variation of scale, rotation and illumination. In this paper, we propose a Content Based Image Retrieval (CBIR) system for HRSI by employing SURF features and boost them through the addition of color, texture and structural information around key points; and by learning a category-specific dictionary for each image class. An extensive experimental evaluation on the well- known UC Merced dataset has been performed and compared with other feature extraction methods including Convolutional Neural Networks. It is demonstrated that our method is quite competitive in terms of performance.