Image Registration By A Statistical Method

An automatic method for accurate registration of digital image data is described. The method uses statistically significant correlations between different data sets to generate control points. Under the assumption that the data can be coarsely registered via sensor geometry models or a few manually extracted ground control points, one can use a statistical technique to "fine-tune" the registration. The technique is based on data whitening, short-space correlations, a false-alarm threshold test, and coordinate remapping. Data whitening permits subsequent analysis of processed images to be based on a Gaussian white noise assumption. A quantitative threshold can then be evaluated which is exceeded only when the corresponding subimages are correlated. This statistical approach often yields a dense uniform distribution of control points. Except for the coarse registration step, the procedure is totally automatic, alleviating the tedious task of manual control point selection. The accuracy of the technique is demonstrated by aligning digital stereo imagery and SAR to simulated Thematic Mapper imagery.