Affine Invariant Feature And Its Application To Target Recognition In Remote Sensing Images

Affine invariant feature is very important to recognition of remote sensing targets. The relationship between the coordination of two images photographed by imaging systems, such as satellite remote sensing , approximately satisfies affine transformation in some certain condition. The geometrical relationship between local regions in two images which are photographed of different views in even close quarter almost satisfies affine transform. Affine invariant feature is the invariance of object under affine transform. Because it overcomes the disturbance of different views, affine invariant feature is the powerful tool of target recognition in different views.First, this thesis introduces the conception of image invariance and summarizes methods of Hu Moments and Affine Invariant Moments based on mathematical models of imaging and affine transform. Then, it studies a new affine invariant feature named Multi-Scale Autoconvolution (MSA) and presents a novel method of local affine invariant feature (Local MSA, LMSA) extraction on the ground of MSA. Finally it puts MSA and LMSA to use in recognition of remote sensing targets.Local affine invariant feature is beneficial to recognition of targets in complex background and partly occlusion. The extraction of local invariant feature includes two parts: Detector and Descriptor. The method of LMSA feature extraction firstly takes the geometrical center of image regions based on certain intensity range in the digital image as Key Points(Detectors), secondly extracts local affine invariant region, and finally the MSA values of these regions are calculated as Descriptors. LMSA feature is made up of Descriptors as a vector set. This feature shows good performance in stability to the affine transformation, viewpoint changing, partly occlusion and noise disturbance.MSA and LMSA feature are suitable for target recognition in different conditions. This thesis presents two recognition methods of remote sensing targets using MSA and LMSA feature, one for small targets, the other for large ones. Real remote sensing images are experimentized, which shows MSA and LMSA feature are efficient in target recognition.