Transformation-invariant extraction of multi-location image features from remote sensing imagery

A novel method of transformation-invariant feature extraction called multi-location saliency pattern is proposed in this paper for object recognition and image matching. Multi-location image features are extracted in salient image points, which indicate image locations with high intensity contrast, region homogeneity and shape saliency. Three distinctive types of fragment descriptors are extracted to form the descriptor vector: pose, regional shape, and intensity (texture) descriptors. Pose characteristics and regional shape descriptors are made invariant to image similarity transformations.