Super-Resolution Reconstruction and Its Application Based on Multilevel Main Structure and Detail Boosting
Round 1
Reviewer 1 Report
This manuscript proposed a new single-image superresolution algorithm based on multilevel main structure extraction and subsequent detail boosting. The manuscript is written clearly, describing both the background and algorithm well, plus includes a comprehensive set of results.
My criticisms are generally minor. The introduction presents single-image SR as the subject of the paper in a weird way. It says that it is difficult to obtain sup-pixel registration. While this may be true, single-image SR is valuable in its own right because of other considerations including, for example, the cost to collect multiple images. It's a legitimate problem that does not need to be justified as the simpler path to purse.
I suggest that the authors name their method to help refer to it in the paper and after it's published.
Also, a brief mention of algorithm run times would be helpful to the reader.
Here are some specific suggestions or questions for the authors to consider:
Abstract: The phrase "main structure" is used frequently in the abstract. Briefly define this phrase for the reader.
Line 69: The 3 images are clearly not acquired at the exact same time due to the building lean. Perhaps this could be refined by saying that the images were captured seconds apart in the same satellite overpass.
Line 187: Instead of the column vectors of S and I, it would be clearer to say that the images are vectorized/reshaped into column vectors.
Line 192: Change to "the values of u and w are straightforward to calculate..."
Line 247: "both the sigmoid function itself and its inverse function have the inverse function" - Is this written correctly?
Equation 12: It looks like 101g here. Write out log and insert some space before it.
Line 303: There are 4 WorldView satellites that are all different. Which one was used?
Figure 4 or 5: It would be good to see the corresponding inset images at standard resolution so that the reader can see the improvement.
Figure 6: In the caption, label which images (top or bottom) correspond to which method. I also recommend listing the top one first (bicubic)
Author Response
We are very grateful for your comments for the manuscript. According with your advice, we amended the relevant part in manuscript. The details of the modifications are in PDF
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper describes a single-image super-resolution method for remote sensing applications. The reconstruction method is based on multilevel main structure and detail boosting. The method is composed by three main steps: 1) The multilevel main structure is obtained based on the decomposition of the image by using a relative total variation model. In this step, multilevel texture detail information is obtained. 2) The multilevel main structure and texture details are reconstructed separately. A detail-boosting function is used to incorporate high-frequency details in the image. 3) The high-resolution final image is obtained by fusing the multilevel structure and the texture information.
Taking into consideration that the proposed method uses a single-image super-resolution approach, the results seems to be quite good, indicating that the method is able to recover some amount of the original missing information.
My main concern is related with the experimental part of the presented work, where some relevant experiments are missing to fully demonstrate the viability of the proposed method. In any case, I will present my reviews following the same structure of the paper.
The claim that “SR reconstruction is a fast and low-cost technology” should be handle with care. Some SR methods are iterative and use to take a long time. Additionally, to reach real time performance, it is necessary to largely increase the amount of physical memory and the number of parallel processing, increasing the overall cost of the SR system. I am afraid that such claim cannot be generalized to all the SR methods.
In page 3, section 2, at the beginning of the third paragraph (lines 132-133), it is said that “In summary, a method of improving the visibility and texture details of a reconstructed image from a HR remote-sensing image with clear edge structure and rich texture details based on the multilevel main structure and detail boosting is presented in this paper”. It seems that instead of HR, it should read LR, as the method is based on the attainment of a HR image from a Low-Resolution (LR) image. Please check it.
The idea of multilevel decomposition resembles the approach of wavelets. Although there are several papers using the wavelet transform for SR, such idea is not mentioned in this paper. Please, conduct a brief discussion about the main differences with wavelets multi-level decomposition.
It is very difficult to read the text superimposed onto the images of Figure 2. Also the text font of the vertical text seems to be wrong.
Indicate the size of the window used in equation (2). When in line 13 it is said that delta “controls the size of the window in Eq. (2)” it is not clear whether it is the size of the window or it modulates the size of the window. Please clarify.
It is very important to indicate the size and the number of bands of the images in section 4.1, Figure 3. Section 4.1 is very relevant as it is the only way to check that the proposed method is increasing the detailed information of the low-resolution images. Please, indicate the number of layers (or levels) used in the algorithm to decompose the images. Provide also the same results than in Table 1 but for different scale factors (not only for scale factor 2).
Indicate the type of subsampling (sampling matrix H) that has been used to decrease the size of the original images. Provide the same results than in Table 1 but for different amounts of noise.
In line 317-318 it is said that “From the subjective perspective, the edge structure is blurry, the sawing phenomenon is also very serious in figure3(a)”, but it seems that it should read “in figure 3(b)”, as (a) is the original image. Please check it.
In Table 2, in the column “Figure”, numbers 4b, 4c and 4d are missing, whereas 4g, 4h and 4i does not exist in Figure 4. Please check it.
It is not understood why methods IBP, VDSR and HE have not been used in the analysis of section 4.1. Please include such algorithms in the evaluation using synthetic images.
In line 339-340 it is said that “The first six groups of images were panchromatic images and the last three groups were multispectral images”, but there are only six images. Probably it should read: “The first three groups of images were panchromatic images and the last three groups were multispectral images”. Please check it.
In the case of multispectral images, please indicate which bands (or combination or bands) have been used to apply the SR process, or if it has been applied to each band independently.
In lines 385-386 it is said that “For the objective evaluation indicators used in this paper, the higher the objective evaluation index, the better the quality of the reconstructed image”. I think that this sentence should be clarified. For instance, the entropy represents the degree of uniform distribution of any energy in space, but a more or less uniform distribution of the energy does not necessary imply an increase of the image resolution. Please reword this sentence.
IBP and HE methods are not included in Table 4, hindering a regular analysis of the experiments. Additionally, it is said in lines 410-411 that “Based on the results of the state-of-the-art SR methods based on deep learning, a training model cannot used for all scales”. Clarify why the deep-learning methods have not been trained for each different scale, i.e. one by one.
For real implementations, it is very important to know the complexity of the proposed algorithm compared with the alternative ones. Please, provide at least the execution times of the proposed algorithm compared with the alternative ones (bicubic, IBP, SRCNN, VDSR and HE) running in the same hardware platform. For those algorithms based on a training process, please include separately the training time and the number of images used for the training process.
Author Response
We are very grateful for your comments for the manuscript. According with your advice, we amended the relevant part in manuscript. The details of the modifications are in PDF
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
I would like to thanks the authors for properly addressing all my comments. I think that the quality of the paper has been improved and it is ready to be published on its present form. Just an English review would be advisable, but the content of the paper is excellent. I would like to congratulate the authors for the very interesting work that they have carried out and presented here.
