Aerial Imagery Feature Evaluation

What features work best for classifying aerial imagery?

Collaborator: Anil Cheriyadat
Sponsor: U.S. Department of Energy Science Undergraduate Laboratory Internships (SULI) at Oak Ridge National Lab

Project Description

High-resolution aerial images are becoming more readily available, which drives the demand for robust, intelligent and efficient systems to process increasingly large amounts of image data. However, automated image interpretation still remains a challenging problem. Robust techniques to extract and represent features to uniquely characterize various aerial scene categories is key for automated image analysis. In this work we examined the role of spatial features to uniquely characterize various aerial scene categories. We studied low-level features such as colors, edge orientations, and textures, and examined their local spatial arrangements. We computed correlograms representing the spatial correlation of features at various distances, then measured the distance between correlograms to identify similar scenes. We evaluated the proposed technique on several aerial image databases containing challenging aerial scene categories. We report detailed evaluation of various low-level features by quantitatively measuring accuracy and parameter sensitivity. To demonstrate the feature performance, we present a simple query- based aerial scene retrieval system.





T. Swearingen and A. Cheriyadat, "Spatial feature evaluation for aerial scene analysis," Proc. of 2012 IEEE Applied Imagery Pattern Recognition Workshop (AIPR 2012), (Washington, D.C., USA), October 2012.