Future work is required to perform more extensive testing, and to incorporate some quantitative assessment. However, from the results above we can form some initial conclusions. First, the Normal model for approximating the noise intensity in combination with the noise estimation procedure works adequately, although the results tended to look speckly. One reason for this is that the method performs the thresholding on a local pixel basis, in contrast to most of the other methods which operate over windows or regions. The speckle could be reduced either by modifying the threshold, but more probably post-processing such as erosion would be simpler and more reliable.
The two techniques we considered for comparing intensity distributions, the Kolmogorov-Smirnov and Cramér-von Mises tests, did not do well. A further disadvantage is that the results are fairly sensitive to the window size which needs to be specified as an additional parameter. The correlation methods for comparing windows, Pearson's rank correlation coefficient and Spearman's rank correlation, also performed poorly. The Kolmogorov-Smirnov and Pearson methods appeared to be particularly sensitive to noise, especially in regions with small dynamic ranges of intensities.
Most promising were the spatial methods. Both the Poisson noise model and the stable Euler number reliably gave good results, which for all the images either matched or bettered the Normal noise intensity model results. However, further tests are necessary to determine the robustness of the peak finding and corner finding procedures used by the two methods.
Paul L Rosin