"... In PAGE 8: ... 4. The values of FFve combined invariants from Section 3 were computed for each image (see Table1 ). We can see that the values of the invariants are fairly stable with respect to image blurring.... ..."

"... In PAGE 9: ... 2: original image, image blurred by horizontal motion (v t = 50 pixels) and image blurred by motion along a diagonal (v t = 30 pixels). Table2 illustrates the property of invariance of these features. Figure 2: (a) original image, (b) horizontal motion blur (v t = 50 pixels), (c) diagonal motion blur (v t = 30... ..."

"... In PAGE 15: ... The performance of GCV was also compared to ML for this problem with the same initial conditions and search procedure for both criteria. Table2 summarizes the results. In every case, GCV achieved a more accurate estimate of the actual blur radius than did ML.... ..."

"... In PAGE 7: ... The stopping criteria that were used for each dataset and the number of attributes in the best subset of features across entropy measures are listed below (see Table 4). Table5 displays the difference between the Classification Accuracy with Blurring minus the Classification Accuracy when there was no blurring. It 0-7695-0981-9/01 $10.... In PAGE 8: ... No reduced data sets were returned for the Car data set. For the Soybean data set, Proceedings of the 34th Hawaii International Conference on System Sciences - 2001 there were examples of improvements as a result of feature selection (the positive numbers in Table5 ), this state was far from universal. Thus our results do not support the observation in [12] that an advantage of blurring-based feature selection is large classification accuracy.... ..."

"... In PAGE 21: ...3 Results The statistics are split into two tables. The rst one sums-up the results obtained by the four methods, for Blurred Image 1 (see Table 1), and the second one, for Blurred Image 2 (see Table2 ). However, we have to keep in mind that neither the FCNR nor the variational methods aim to minimize these errors (the FCNR controls the structure of the noise, and the variational method controls the regularity of the image).... In PAGE 21: ... Looking at the di erent zones in detail, we observe that the FCNR apos;s are more e cient than the variational method on Zones 1 and 2 and less on Zones 3 and 4, which con rms the fact that the Total Variation penalizes textures but preserves strong edges (it does not generate ringing in there vicinity). If we turn now to Table2 , both wavelet-packets based algorithms yield poor results. The convolution operator of Blurred Image 2 is actually not invertible so that we left the framework in which these methods are valid.... ..."

"... In PAGE 3: ...light visual change on the images (see Fig. 1). We used the non-smoothed image as a query to nd its smoothed coun- terpart in the set of twenty smoothed images. The retrieval results of this experiment are given in Table1 . The unrelia- bility of the color ratios p and m under blur is apparent: only for a few of the queries the relevant image was found with rank 1.... ..."

"... In PAGE 21: ...3 Results The statistics are split into two tables. The rst one sums-up the results obtained by the four methods, for Blurred Image 1 (see Table1 ), and the second one, for Blurred Image 2 (see Table 2). However, we have to keep in mind that neither the FCNR nor the variational methods aim to minimize these errors (the FCNR controls the structure of the noise, and the variational method controls the regularity of the image).... In PAGE 21: ... Let us indicate also that the parameters of these methods has not been xed with regard to these statistics but on visual criterion. In regard to Table1 , note that the invertible FCNR and the variational method yield, on the whole, comparable statistics while the results are a little worth for the non-invertible FCNR, and much worth, on any region, for the wiener lter. Looking at the di erent zones in detail, we observe that the FCNR apos;s are more e cient than the variational method on Zones 1 and 2 and less on Zones 3 and 4, which con rms the fact that the Total Variation penalizes textures but preserves strong edges (it does not generate ringing in there vicinity).... ..."