An Upscaling–Downscaling Optimal Seamline Detection Algorithm for Very Large Remote Sensing Image Mosaicking
Round 1
Reviewer 1 Report (Previous Reviewer 1)
Perhaps I have pointed this out before, but please add an explanation for Figure 12, which is missing from the body text.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report (Previous Reviewer 2)
The quality of the modified manuscript has been greatly improved and its topic is very important for remote sensing image mosaicking. I have no further question about the manuscript, so accept is suggested.
Author Response
Thank you for reviewing our manuscript again.This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
This method proposes a two-stage graph cut method for seamline detection and composition of high-resolution images.
Experimental results show that the proposed method is the fastest and the accuracy is comparable to native graph cut.
The description of all formulas is very unclear. It is difficult to tell what the algorithm is at present, unless one already understands the algorithm for graph cutting of images.
For example, C_{NA/NB} is a scalar value in the formula definition, but the input is an RGB-IR or RGB vector value. There is no definition for x1,x2,y1,y2,xp,yp, and x,y without subscripts appear at the same time.
The entire \xi matrix of pixel values and the scalars (or vector values) Ip and Iq of pixel values are in the same formula.
W in equation (7) has no definition.
Equation (9) has no definition for almost all variables.
The key idea is reducing the number of nodes in the graph cut in a course to fine search, however, the hierarchical method is a typical image processing technique and is not particularly new.
It should be made clear how there is a problem that cannot be dealt with in the usual way for the problem this paper is trying to solve.
I think there is value in this paper, which tries to achieve a very large resolution image, so I hope it can be improved to the point where it can be readable as a paper.
Other comments:
1. Please describe problem setting: registration between two images are given?
2. Please define “resample+ Dijkstra” method in detail.
3. Programming language and library information is required for experiments
4. Why resize factor is 10. The other case comparison is requires.
Minor:
Fig.1 calculate geographic overlap area”r”
apply?
we receive? That
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
To reduce the amount of calculation for finding seam lines in the process of image mosaicking, the paper propose a stepwise strategy for getting pixel-level optimal
stitching lines for large remote sensing images through an upscaling-downscaling image sampling procedure. The graph cut algorithm is applied to find an energy-optimal seam-line in the reduced image and nested within identified stripe to seek the pixel-level optimal seam-line of the original image. Compared to the existing algorithms, the proposed method has fewer spectral differences between stitching lines and less crossed features in experiments and consumes less than 10 percent of time compare to SLIC+Graph cut method.
I have read this paper and the comments are as follows:
1. Line numbers are missing in the manuscript. Some typos exit in the last line of the first paragraph in the second of chapter two.
2. Some of the images in Figure 1 are of low resolution, and their typographical structure needs to be optimized.
3. Some typos occur in Section 2.2.2,and some of the text in Figure 2 is blurred.
4. Mathematical formulas should be in the middle, some place should be modified.
5. Compared with the text of the manuscript, Figure 3 and Figure 5 are a little too large. Figure 4 is blurred.
6. Some values in Table 2 are missing. It is recommended to complete them. Please explain clearly if they cannot be added.
7. References of the last two years should be supplemented.
Over all, the topic is very interesting and has high application value, the method proposed in the manuscript is advanced and effective, so minor revision is suggested.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The submitted manuscript is proposing a seamline algorithm to merge more satellite images in a single map. The algorithm seems theoretically well-posed and the subject is definitely interesting for remote sensing, in particular for the automatic image segmentation. In this manner, an object can be detected in a single image, without the problem of having it divided in two or more images.
However, I have a question. How is the algorithm depending on the latitude of the places recorded by satellite images? At high latitudes, the deformation is larger, is the method suitable also for them? Please provide in the manuscript some examples of the method applied to satellite images of regions at high latitude such as at low latitude.
Author Response
Please see the attchment.
Author Response File: Author Response.docx
Reviewer 4 Report
To address the problem that Graph Cut is slow or failing in extracting optimal seams for large overlapping regions, the authors propose a strategy of using Graph Cut twice. They obtain a coarse boundary by searching on the downsampled graph, and then search again in the vicinity of the coarse boundary. The article illustrates the improvement in efficiency of the algorithm and its ability to process for large images through a mosaic of four pairs of images.
In terms of innovation, the novelty of this work is limited. The idea of hierarchical processing from coarse to fine for large images is a commonly used engineering idea, but academic contributions are scarce.
To demonstrate the effectiveness, the paper evaluates the color difference between the two sides of the optimal seam to illustrate the advantages of the newly proposed algorithm. However, from a perspective of initial motivation, Graph Cut targets the minimum color difference while the superpixel preprocessing algorithms are designed to maintain structural integrity. Therefore, it is not convincing only with the color difference metric to illustrate the effect.
The article has a large number of writing errors, as shown in grammar, vocabulary, and punctuation. For example, Figure 1 is not easily understood due to inaccurate words. Some typical errors are listed below.
l Full name of SLIC was not given.
l Ma and Sun [13] developed a novel A* algorithm [14] to obtain the global optimal splicing line by optimizing the improved(feels like not finished)
l Finally, apply? color leveling and mosaic treatment according to this optimal seam-line.
l Notations are not explained for Eq. (7)(9)(10).
Author Response
Please see the attchment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Again, please review the expression of the equation. There are too many undefined variables.
The concept itself is understandable and can be outlined, of course, but the detailed method is not defined.
1．Define C in vertex V. Multiband, right? Vector? Or the average of each spectrum?
2. To define R_i, we need to define C_NA_i to know.
3. C_NA of B(L) is undefined, and if it is a vector, then B is a vector value. But since R is a scalar, perhaps B is a scalar. Given that, the definition is undefined.
4. Once again, there is no definition of \xi. It appears to be a pixel position or an entire image. The definition of graphs and the definition of energy have different formulas, which makes them difficult to understand.
5. Equation 7 is not known because the variables, p, q, I, and function D are not defined.
6. Since \xi is not defined, we do not know if E_data(\xi) and E_data(I_p) or E_smooth(x,y) and E(\xi) can be substituted.
7. In Sec. 2.4, please define S_k, T_k, k, and K.
8. Please define h and j. Are they hyperparameters?
9 Please define P
10. What is \alpha and how to use it? I cannot understand this section of 2.4.
11. In the experiment, resample+ Dijkstra is not well defined. How do you get a full-resolution seamline from resampled one?
12. In Table 2, please define the average color difference, and number of features crossed.
13. In Table 2, what is refinement of the color difference range interval? and what is cost ? what is the percentage?
14. In addition, Figure 12 is completely incomprehensible because there is no explanation.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The authors have properly answered my observation.
Author Response
The authors would like to express their appreciation for your time and effort in reviewing the manuscript again as a reviewer.
