Towards the validation of the method,we used images of unknown noise models and have compared our method with well known filtering techniques such as the bilateral filter and the Non Local Mean approach.Some qualitative results and comparisons are shown in [Fig. (4 and 5)]. As for quantitative comparaison, number of methods have been used in the literature for such comparisons,given the absence of knowledge on the noise model we have considered calibration patterns and studied the behavior of the tested approaches on these patterns when observed from different digital cameras. We focus on the noise reduction on these patterns which is equivalent to the reduction of the standard deviation relative to a uniform patch. . Table shows the performance of each filtering technique in terms of noise reduction for different digital camera models. Considering this criteria, results show that our method outperforms the other techniques. This is explained by the fact that in absence of texture or structure the Random walk acts as an isotropic filtering. The bilateral filter consider the pixel location while denoising which limit the influence of distant pixel in the filtering process. The non local mean,uses to compute neighborhood similarity the L2 distance which is more sensitive to outliers then the L1 distance we use..
As for qualitative results,we selected for each method the most suitable parameters that gives the best compromise between noise reduction and detail preserving. If we consider the method noise images which corresponds to the difference between the original and the filtered image (see fig 4), we notice that the random Walk filtering produces better results than the bilateral filtering in terms of texture and small details preserving. This is explained by the fact that in our approach
we make a structure tracking and we integrate information about image statistics in the model. The NL mean technique achieves the best results in term of small detail preserving since its method noise contains less image information then the two other techniques. This is due to the fact that the NLmean algorithm scans all image pixels to select the best candidate while denoising a given pixel and this make it very slow in terms of computation time.
帮帮忙,谢绝机译