Registration is an important aspect of any image processing system. Registration may be interpreted as automatic determination of local similarity between two structured data sets. Several digital techniques have been used for registration. Principal among these are cross-correlation, normalized cross-correlation and minimum distance criteria. Fast algorithms like sequential similarity detection for translational differences have been developed. In all the above methods, a portion of one image is taken as reference and similar portion in the other image is located by carrying out a search over the other image. There may be translational, rotational and scale differences between the two imageries. Here, in this paper, coordinates of the two imageries have been related by affine transformation and these transformation coefficients are then evaluated by carrying out a search for similarity between the two imageries. In order to carryout this search in an optimized way, an algorithm using sequential simplex method (Box's Method) has been developed and implemented. The method is essentially a gradient type and searches the optimum with the steepest route. The method has been applied to register two Landsat scenes. Although, the method is general, in the illustrative example, only the translational differences have been considered. The results of the same are shown in the plates. Also, because of the presence of noise, there exists the limit on the accuracy of the registration. Here, the two imageries are considered to be the sample functions of homogenous random field with known auto-correlation enunciation Also, the non-similarity between the two imageries is considered as noise in one of the imagery. With these assumptions, the upper bound on the accuracy of the registration has been evaluated.
Available at: http://works.bepress.com/arun-kulkarni/2/