Improving high resolution satellite images retrieval using color component features

Abstract

This paper highlights multi-scale color component features that improve high resolution satellite images retrieval. Color component correlation across image lines and columns is used to define a revised color space. It is designed to take simultaneously both color and neighborhood information. From this new space, color descriptors namely RIULBP (Rotation Invariant Uniform Local Binary Pattern), LV (Local Variance) and a modified version of LV (smoothed LV) are derived through Dual Tree complex wavelet transform (DT-CWT) or scale-invariant features transform (SIFT) representations. The features obtained offer an efficient way to represent both color and texture/structure information.We report an evaluation of the proposed descriptors according to different similarity distances in our CBIR (Content-based image retrieval) schemes.We, also, perform comparison with recent approaches. Experimental results show that color LV descriptor combined to SIFT representation outperforms the other approaches.