A novel multi-scale relative salience feature for remote sensing image analysis

Abstract

This paper presents a novel feature for remote sensing image analysis, called multi-scale relative salience (MsRS) feature. It is constructed by modeling the process of feature value changing with scales. Firstly, the multi-scale observation values at each site are obtained by convolved with recursive Gaussian filters for efficiency. Secondly, the multi-scale observation values are compared with the initial value to generate the relative salience. Lastly, the relative salience between multi-scales are embed into a single feature called the MsRS. The scale in MsRS has explicit spatial meaning which is convenient to choose appropriate scale for specified object. In the MsRS map, the inner of each object become more consistent, while the contrast between object and background is enlarged. The MsRS can be used as preprocessing step of many applications, such as segmentation. Two state-of-art segmentations (the mean shift and the statistical region merging) are taken into experiments and the results proved that it brings improvement obviously.