| With all the development in remote science at home and abroad, there is a leap in both quality and quantity of remote data, which has brought us into the explosion of remote data. Becoming the main type of data used in remote science, high-resolution remote sensing image has abundant spectral information, clear spatial layout and sharp geometrical structure. Recognition of high-resolution remote image scene is the foundation of many remote-technology applications. It is of great significance for military and disaster relief that the recognition of remote sensing scenes can be accomplished automatically. However, there are few researches using the newly developing method for scene representation based on biological vision, semantic image method, in high-resolution remote scene recognition, while BOW method has been applied to solve this problem in some researches.Trying to catch up with the novel research in scene representation, this paper systematically studied the summary statistic mechanism in scene recognition during human’s early vision, based on which a new method for scene representation, summary statistics, was proposed and used to recognize high-resolution remote sensing images. By extracting 4 typical summary features, average orientation message、scene clutter、scene variation and mean size of objects, we built a relatively complete representation of scene’s global properties and simulated the summary statistic mechanism in human vision properly.In order to prove the effectiveness of summary statistics, we did classification experiments on two remote datasets and compared with Gist and BOW, including typical high-resolution remote image scene. The accuracy is relatively higher. We also did experiments to classify non-remote sensing images, which resulted in slightly higher accuracy. The experimental results demonstrated that summary statistic method can represent scene effectively and the method realizes visual system’s function of scene recognition to a certain extent. |
PDF Full Text Request