The authors describe a technique for determining spatially variablewindow sizes for image processing tasks, including restoration, stereomatching, and motion correspondence. It is difficult to achieve a goodtradeoff between smoothness and fine details in many imaging problems.The spatial resolution needs to be adjusted adaptively in order toobtain a good result, and heuristic techniques often fail in the generalcase due to the rich structures found in natural images. This papersuggests a framework for systematically selecting an appropriateresolution adaptively using a statistical technique.
First, the task is formulated as a plausibility test of eachinstance at each pixel location. For example, the number of possibleinstances is equal to the number of possible intensity values for imagerestoration. The test is formulated in such a way that an instance isplausible if its likelihood is larger than the likelihood of its notoccurring. Second, the window size is computed for each instance bygrouping pixel locations where the plausibility test for the instance issatisfied. Third, the task is solved by choosing the instance with thelargest window size at each location.
The idea is interesting. However, the rationale for the third stepabove is not clear, and the authors’ experiments only provide simplesynthetic images for image restoration, which casts some doubt on thetechnique’s ability to resolve finer details.