Illuminant Estimation Using Adaptive Neuro-Fuzzy Inference System
Round 1
Reviewer 1 Report
This manuscript presents in full detail a very complex, largely statistically based, method of assigning the color of the illumination that might have been present when a given image was made. It incorporates methods for combining results obtained by other previously described methods to obtain results closer to the "ground truth" for the test images used. The Methods section is more complete than might be expected for this type of paper and I find no flaws in the individual components. I only have a few questions that I think are worthy of more discussion (some of the methods could be shortened if length is a problem) followed by a few comments on possible textual errors or ambiguities.
Questions:
(1) Why would anyone think of perfoming this particular sequence of operations to solve this problem? The authors provide information on experiments performed to settle on values of some of the parameters (number of clusters in k-means clustering, for example) but a little more rationale for why, for example, two levels of clustering rather than one were used would be welcome.
(2) In Figs. 3-7, it is interesting that in most instances, a small difference of a few degrees in the AE (angle error) makes practically no difference on the impression the image gives on the reader. On the other hand, some of the methods as seen in Figs. 6 and 7 are quite bad even with similarly small AE. One would like an explanation for the lack of correlation between AE and visual impression. Is AE really a good measure of error?
(3) Fig. 8 is very confusing. I think we are seeing the distribution of the minimum AE across the 8 methods tested individually with the conclusion that there is at least one method with an error less than the abscissa for the number of images given by the corresponding ordinate. This figure provides no information on how many of the 8 methods give the minimum AEs plotted. That being the case, this figure does not support the conclusion (l. 367) that "This figure shows that if we use an appropriate, linear or nonlinear, combination of these unitary algorithms, it is possible to achieve better accuracy of illumnation estimation." Figure should compare minimum AE distribution for each method if used alone (different colored curve? different plot?) versus minimum AE obtainable with chosen combination of methods.
(4) In Table 2 it is clear that at least a few of the learning-based methods outperform the method proposed here. The authors suggest (l. 304-305) that their method might nonetheless be preferred for "ease [of] encoding into image signal processor". The thought occurs to this reviewer that what this paper really shows is that learning-based methods can outperform even the most carefully designed and tested combinations of statistical and fuzzy inference systems. Perhaps this should be mentioned in the Abstract. Any response?
Comments:
--In ll. 156-157, present tense should be used instead of future tense as authors are describing the method, not future experiments.
--The meaning of the colors of the curves in Fig. 2 is not described.
--The term "unseen image" (l. 220, section 3.4) probably means "image not included in the training set"?
--The abbreviation "ISP" (l. 305, section 4.2) is an unfortunate choice in view of the common use of this abbreviation to mean "internet service provider".
--It is not clear in Tables 2 and 3 whether training was done with both data sets combined into one large set for training and then tested separately, or were the two data sets trained and tested separately? Training on one set and testing on the other would give an even better picture of the generality of the method.
--In l. 327 "conduct" should be "conducted".
--In Table 7, why does the line k_1 = 3, k_2 = 1 appear twice with different statistics?
Author Response
Dear Editors and Reviewers,
Please see the attachment.
Sincerely yours,
Luo Yunhui
Author Response File: 
Author Response.pdf
Reviewer 2 Report
In this paper, the authors have proposed an ANFIS based approach for illuminant estimation.
Overall, the paper is very well written and the reviewer commends the authors.
The authors have provided an adequate introduction, background and references. The motivation and contribution have been presented. The proposed method is clearly explained step by step with equations and illustrations. The proposed approach seems to provide good results based on the comparative statistical metrics presented by the authors. The authors have also analyzed the impact of some of the factors influencing the performance of the proposed method. In addition, the practical applicability was also discussed by providing the implementation details including the test data set, the configuration of the computer system used, the computational time for implementing the proposed approach on the test data set, etc.
Reviewer comments and recommendations:
- The Github link shared by the authors on Pg 10 (https://github.com/yunhuiluo/AnfisIllests) for showing their Matlab implementation was not accessible by the reviewer.
It looks like there was a typo error with an extra ‘s’ at the end of the link (https://github.com/yunhuiluo/AnfisIllest).
The authors are requested to check and update the link presented in the paper. - The reviewer recommends adding a list of abbreviations at the beginning or end of the paper as an easy reference for the readers.
Author Response
Dear Editors and Reviewers,
Please see the attachment.
Sincerely yours,
Luo Yunhui
Author Response File: Author Response.pdf
