Analytic Methods of Image Registration: Displacement Estimation and Resampling

摘要

Registration algorithms are developed and evaluated by using oversampled scanningimagery directly and staring imagery with a subpixel autocorrelation model. A Fourier transform-based method, Phase Correlation, is enhanced to remove edge effects, to accommodate nonintegral shifts, and to resolve ambiguity in the interpretation of its output. The result is shown to be accurate over a wide range of misregistration and to be capable of detecting cloud parallax and other relative-motion effects. A family of gradient-based methods is also derived and is shown to include older methods: the Image Displacement Estimation Algorithm (IDEA) and the Gradient Estimation Method (GEMS). One member of the family, the Canonical Gradient Estimate (CAGRE), proves to be generally superior, as long as the noise-to-clutter ratio is not unusually large compared to typical values for Earth backgrounds. Resampling methods are also tested: linear, spline, phase-shifting, and cubic convolution; cubic convolution performs marginally better for all values of shift. The net result here is a substantial improvement in the state of the art of analytic image registration, which may permit the use of frame differencing as a moving-target indicator in the most cluttered natural backgrounds. (Author)