The measures of dissimilarity D(A, B) between arbitrary gray-scale images A and B investigated in this paper are useful for the comparison of algorithms devoted to the same task (e.g. edge-preserving noise smoothing, restoration, edge detection, thresholding) and for evaluating the amount of image change caused by an affine transformation. There are several papers describing measures of correspondence or similarity between two binary images, and a few of them extend their scope to the quantitative comparison of isolated objects of gray-scale images. Although the set of dissimilarity measures proposed in this paper is applicable to arbitrary gray-scale images, for practical purposes this measure is performed only between very similar images as in the cases mentioned above. First of all, we consider the requirements posed to D. Then we propose a multi-stage dissimilarity measure, in which each stage (pixel-to-pixel, pixel-to-window, window-to-window, and image-to-image) can be based upon different distance measures, thus originating several variants of D. Some properties of these variants, related to the requirements posed to D and to the structure of the compared images, are examined and made the object of experimental verification. Numerous experimental results, obtained with real-world images, illustrate the performance of the proposed measure in relation to the above-mentioned tasks.