Blind Deblurring Using Internal Patch Recurrence: Supplamentary Material

1. We discuss the mapping between the error-ratio measure computed with the nonblind deblurring algorithm of Levin et al. [11] and the error-ratio measure computed with the nonblind-deblurring method of Zoran and Weiss [22]. 2. We explain why down-sampling an image using a sinc kernel has an effect of aliasing-aware sharpening. The error-ratio measure r (Eq. (15) in the paper), which is standardly used to quantify the performance of blind-deblurring methods, depends on the type of non-blind deblurring algorithm used for the final deblurring stage. The error-ratios reported in our paper (and also used in [18]) were computed with the current state-of-the-art non-blind deblurring method of Zoran and Weiss [22]. Previous studies (e.g., [12,13]) reported error-ratios computed with the non-blind deblurring method of Levin et al. [11] (the state-of-the-art at that time). In the figure below we re-plot the cumulative distribution of error-ratios, but this time computed with [11]. As can be seen, the absolute error-ratio values are different than those in the corresponding graph (Fig. 5 in the paper) of error-ratios computed with [22]. Nonetheless, the two graphs reflect the same relative behavior of all the methods. In the table below we report the average performance, worst-case performance and success rate of all algorithms, this time using error-ratios computed with the non-blind deblurring of [11]. We note that for this setting, the blind deblurring paper of Levin et al. [13] reported a threshold of 3 between good and bad visual results. Consequently, we regard success rate in this context as the percent of images with error-ratio smaller than 3. As in the corresponding table (Table 1 in the paper) of error-ratios computed with the non-blind deblurring of [22], this table shows that our method and the method of Sun et al. [18] outperform all other methods in all three categories. The average performance of our method is close to that of Sun et al. [18], while our worst-case performance is significantly better. Note that also when using this measure, only three methods (the same three) attain an average error ratio smaller than 3: Our method, Sun et al. [18], Xu and Jia [19]. The figure below further shows a scatter-plot of both types of error-ratios. Each point in this plot corresponds to a kernel produced by one of the 7 tested blind-deblurring methods on one of the 640 blurry images in the database (once used with the non-blind deblurring …