Lifting Scheme-Based Sparse Density Feature Extraction for Remote Sensing Target Detection

Round 1

Reviewer 1 Report

This reviewer analysed carefully all the changes Authors carried out. The efforts performed by Authors have been deemed effective by this reviewer, therefore my opinion is that the manuscript can be published in the present form.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper presents a method for improving convolutional neural networks in order to make them more suitable for the task of target detection.

I think the paper should be accepted. The paper clearly presents the goal of the research and its motivation. Moreover, the paper contains good literature overview. Experiments presented in the paper are sufficient. Experiments are based on existing test data sets which were not prepared by Authors, however it is correct. The structure of the paper is also correct.

My only concern is such that Figure 6 is too small and it is hard to read. Moreover, I think that Authors should present also another figure or a table showing the level of improvement caused by the algorithm they propose with regard to tested existing algorithms. It can be average values for the ranges used in figure 6.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper describes the application of the lifting scheme to feature extraction from remote sensing images. The relationship between the lifting scheme and convolution is exploited to improve the detection precision by taking the information about adjacent feature points into consideration. Experimental results on the target (ship) detection in both SAR and optical remote sensing images prove the improvement in detection performance and computational efficiency as compared with the widely employed strided convolution. This paper is an extension of the application of the lifting scheme to remote sensing (scene classification), which has been published by the same author group (He et al., Remote Sens. 2019, 11, 2648; doi:10.3390/rs11222648, hereafter He_RS2019). Although the current manuscript provides some valuable insights into target detection in remote sensing, the authors should reconsider the following issues before warranting publication.

(major)

L58 “This paper introduces the lifting scheme [16–18] ...” – please mention here the paper He_RS2019, which is cited as [32]. Since both He_RS2019 and the current manuscript deal with applying the lifting scheme to the data analysis in remote sensing, it is important that the authors explicitly describe the similarity and difference between the two papers here.

L169 “2.3. From Wavelet to Lifting Scheme” – this subsection is nearly the same as Sec. 2.1 of He_RS2019. Similarly, the following subsection 3.1 is akin to Sec. 2.2.1 of He_RS2019. Surely, it is desirable that the description of a paper should, in principle, be self-contained, but such repetitions from the previous publication (without referring to it until L206) are not recommendable. A similar argument also applies to Figs. 2-4, which are based on Fig. 2 of He_RS2019. Thus, my advice is that the authors should reconsider a better and acceptable way to explain the new method by emphasizing the difference from (and similarity with) He_RS2019 since the goals of the two papers are different (scene classification vs. target detection). In this way, the current paper would be valuable for the reader who wishes to try the new methodology.

L197 Fig. 2 and L204 Fig. 3 look identical. Moreover, the same figure is utilized in Fig. 4 as a component. Please remove (or at least reduce) the redundancy, in addition to the proper reference to He_RS2019 as mentioned above.

L206 Please explain why the form of eq.(8) is appropriate as the most simple form of the lifting scheme. Also, how is the initial value is given for the parameter set (h0, h1, h2), and how will they be updated in the back-propagation algorithm?

L287 Please explain why the overall features of curves are different among (a), (b), and (c). For example, the deviations of different curves are larger for (b) than for (a) or (c); Why the values of recall are particularly small in (b), etc.

(minor)

Please consider the following issues and corrections to keep the quality of a published paper.

L10 “Experiment results” – “Experimental results”. The same applies to L306, L386, etc.

L63 “Daubechies and W.Sweldens had proven that ... [22]” – “Daubechies and Sweldens [22] had proven that ...”

L64 Please spell out FIR.

L80 “As shown in literature [15]” – “As shown by Springenberg et al. [15]”. The same applies to L310.

L86 “a high-efficient algorithm for” – “a highly efficient algorithm for”

L88 “The lifting scheme reserves the nonlinearity of the pooling layer thus is ...” – “The lifting scheme reserves the nonlinearity of the pooling layer. Thus, it is ...”

L107 “the existed methods” – “the existing methods”

L142 “AlexNet [30], VGG [2]” – “AlexNet [30], and VGG [2]”: please spell out the abbreviations.

L169, L189, etc. Please put a comma (or a period) after an equation (or a set of equations) since equations are part of a sentence.

L204 “There is one channel” –“There is only one channel”

L204 Please define the “z-transform” precisely.

L241 Please explain why the “ship detection task” is considered to be appropriate for testing the proposed scheme.

L237 Please explain AP_50 that appears in Tables.

L287 Please insert a space between the variable name (such as IT) and its unit (ms). Please explain the meaning of (+2.0) etc. in the Table caption.

L329 “Noted that” – “It is noted that”

L330 “perform” – “performs”

L347 A comma is needed before “respectively”.

L365 “availability” – “applicability”?

L370 “and 0.7% respectively while are” – “and 0.7%, respectively, while they are”

L380 “This paper introduces” – “This paper has introduced”

L383 “the method that extracting ... and then downsampling” – “the method that extracts ... and then downsamples”

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors have responded mostly adequately to each of my comments. For further improving the paper quality, the authors may consider the following points. No further review is mandatory.

L64 “Daubechies and W.Sweldens [22]” – “Daubechies and Sweldens [22]” (In the citation, it is common to cite only the family names of the authors. Otherwise, “I. Daubechies and W. Sweldens” would be more balanced and appropriate.) The same applies at L190.

L74 “Motivated by these facts, we introduce the lifting scheme into ... “ – “Motivated by these facts, in the present paper, we introduce the lifting scheme into ...”

L192-193 “Transformed to the z-domain, the two-channel filter bank representation is equivalent to Eq.(4), with A(z) and D(z) to represent the z-transform of a and d.  <eq. (4)> “ – “Transformed to the z-domain, the two-channel filter bank representation is equivalent to

    <Eq.(4)>,

with A(z) and D(z) representing the z-transform of a and d. “ (In this way, the equation can be properly incorporated in the sentence.)

L199 “the left (right) of the equal sign” – “the left-hand (right-hand) side”

L259 Fig. 3: There seem to be three squares in the ship image given in the middle. It would be nice if the meaning of these squares is briefly explained in the caption. Also, it would be better to highlight the squares since the thin lines are hardly recognizable.

L284 “The higher AP indicates better detection performance” – “The higher AP value indicates the better detection performance.”

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

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

The paper entitled “Lifting Scheme-Based Sparse Density Feature Extraction for Remote Sensing Target Detection” treats about a topic of the highest interest in image classification and target detection scope. Therefore, the topic clearly falls under the journal scope. The declared objective of the manuscript is to introduce a novel method for target detection, namely the lifting scheme capable to lower the computational complexity and on the other hand to improve the detection precision.

Concerning the manuscript, the typographical outline is nearly satisfactory. The used language is fluent apart some typos and some repetitions spread over the manuscript (see the section specific comments for some examples). Abstract and Introduction introduce the reader into the treated topic finely, clarifying very well the aims of the manuscript. Keywords are pertinent to paper content and appropriate. The highlights are missing. Materials and Methods are good. The References section is rich. Graphic representations are very fine.

Entering in the very merit of the paper is opinion of this reviewer that the manuscript is well written and the experimental part is well conducted. The evidences seem to confirm the hopes of Authors concerning the better performances of the proposed method.

In conclusion, this reviewer recommends considering the paper after “minor revisions”.

Specific comments

Lines 17 – 18: Please, change “Histogram of oriented (HOG) feature” into “Histogram of oriented gradient (HOG) feature”.

Line 24: “precision of predicting” or “accuracy of predicting”? Please, clarify.

Lines 72 and following appear to be the same of lines 81 and following. Please avoid repetitions.

Line 113: Please, change “necessary to utilized” into “necessary to utilize”.

Line 116: this is substantially a repetition of line 14. Please, avoid repetitions.

Line 155: After the sentence “W. Sweldens proposed” the reference is missing.

Figure 2(a): it appears that figure 2(a) is a repetition figure 1.

Lines 157 – 161: it appears that this section is devoted at the description of figure 2 (b). However, the three described steps do not fully correspond to the letters used in the figure. Please, improve the section using the same parameters introduced in figure 2(b).

Lines 162 – 163: in the sentence “Daubechies and W. Sweldens have proved” the reference is missing.