A.G. Tashlinskii, I.N. Kaveev, S.V. Voronov
One of the main problem in processing of digital images, which defined by discreet sampling grid, is image registration, which consists in match making between conjugate points of two or more images. Data images may have global and local spatial distortions. Image registration in conditions of low noise is well-studied. However in cases of intensive noise, correlated pulse noise, like clouds, in particular, it requires to develop new reliable registration procedures. Based on pseudogradient adaptation proposed method consists of following stages: set of reference fragment defining, initial approximation optimization of pseudogradient procedures for image registration (location) parameters of reference fragments estimation; fragments registration parameters disruptions culling; global registration model parameters estimation using matrix of fragments registration parameters estimations; filtering and interpolation of fragments registration parameters estimations; prediction of registration field between data fragments centers; estimation of resulting registration field reliability. Proposed method let us to perform image registration in conditions of intensive pulse noise. The method is intended to registration mathematical model parameters identification, computational resources requirements reducing and estimation of forming registration field reliability. To rise the accuracy of image registration the techniques of disruptions (output estimations vectors for a given confidence interval) culling of pseudogradient fragments locations estimation. For the simulated images at a signal/noise ratio of two, and affine models of the standard deviation the registration error was 0,8 – 1,1 pixel, for real satellite images – 0,9 – 1,6 pixel.