Department of Image Processing

Abstract:

Positional accuracy of an area in image is dealt with. Two main sources of uncertainty that influences the positional accuracy are considered: 1. uncertainty of classification, 2. precision of image registration. It is assumed that several spatial areas were recognized as a result of Bayesian classification of a given digital image. It means that these objects are not determined precisely as crisp objects. Registration of the given image is supposed to be determined by a linear transformation, coefficients of which were estimated from ground control points by means of Bayesian approach. This transformation is therefore imprecise as well. It will be shown how to properly combine these two partial imprecise results to get the overall spatial precision of each area in the given image. Each resulting individual area is expressed by its boundary which is approximated by a polygon with imprecise vertices. Precision of the vertices is computed from the initial data by means of Bayesian estimation so that any border point between adjacent objects is unique though imprecise. Special probability distribution (close to Gaussian) was designed to model the positional accuracy of the vertices.