An Aircraft Object Detection Algorithm Based on Small Samples in Optical Remote Sensing Image
, Changqing Cao 1, Xiaodong Zeng 1, Zhejun Feng 1, Jingshi Shen 2, Weiming Li 2, Bo Wang 1, Yuedong Zhou 1 and Xu Yan 1Round 1
Reviewer 1 Report
This paper aim is to propose the aircraft detection algorithm, which can detect aircraft objects with small sample.
Tis kind of algorithm is very useful since research on aircraft object detection based on
optical remote sensing images is strategic for military object detection and recognition
and since to obtain spaceborne optical remote sensing images is costly and difficult.
Experiments show that the proposed method is simple and effective for detecting aircraft objects. Furthermore, this kind of research has high value for improving counterattack capability.
The study is interesting, well structured and well written. However, it could be further enhanced, focusing more attention on the conclusions.
Furthermore, with regards to the literature, it should be broaden. In particular, I suggest to consider the following works:
https://www.researchgate.net/publication/328266643_Cluster_Analysis_An_Application_to_a_Real_Mixed-Type_Data_Set
I encourage the authors to refine their paper to make it available for publication in the journal.
Author Response
Date: August 3, 2020
Dear sir/madam,
Thank you for your review and comments on our paper applsci-890551. We have revised the manuscript according to your kind advice and detailed suggestions, and have highlighted the revised part. The point-by-point answers to the suggestions are listed as below.
Thank you very much for all your help, and we are looking forward to hearing from you soon.
Comments and Suggestions for Authors:
This paper aim is to propose the aircraft detection algorithm, which can detect aircraft objects with small sample.
This kind of algorithm is very useful since research on aircraft object detection based on optical remote sensing images is strategic for military object detection and recognition and since to obtain spaceborne optical remote sensing images is costly and difficult.
Experiments show that the proposed method is simple and effective for detecting aircraft objects. Furthermore, this kind of research has high value for improving counterattack capability.
Point 1:The study is interesting, well structured and well written. However, it could be further enhanced, focusing more attention on the conclusions.
Response 1:Thank you for your question.
We have revised the conclusion.
Related information is added on lines 260-263 of the revised manuscript.
Point 2: Furthermore, with regards to the literature, it should be broaden. In particular, I suggest to consider the following works:
https://www.researchgate.net/publication/328266643_Cluster_Analysis_An_Application_to_a_Real_Mixed-Type_Data_Set
I encourage the authors to refine their paper to make it available for publication in the journal.
Response 2:Thank you for your question.
We have modified the the paper and added the literature.
Related information is added on lines 179-181 and 357-359 of the revised manuscript.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors are presenting a paper discussing object detection in the field of computer vision, and particularly applied to the detection of aircrafts. The manuscript deals with object detection algorithms and its use in different fields (as stated in the introduction). The case is that the paper is very concentrated in aircraft detection, but ignores the use of these techniques in other research areas that are not properly cited. At least in the introduction should be explained.
Just to put an example, the authors are not citing any work related with the detection of asteroids, meteors and other atmospheric events. These fields have used computer vision techniques and detection algorithms as well and deserve to be cited (Trigo-Rodríguez et al., 2008; Grav et al., 2011; Soula et al. 2014).
The introduction fails in explaining to a general reader what is an object detection algorithm. For example, a diagram could be useful (Raghunandan et al., 2018). In general the paper is not providing a case example either.
They also fail to demonstrate how significant is the progress made by their initiative in comparison with other algorithms or techniques. On the other hand, it is unclear the exact contribution and novelty of this research to the field. You should put your results in an overall context. I also miss a general paragraph at the end of the introduction with the main goal of the paper.
Section 2.2.1 Try to explain what is your contribution to the mean shift algorithm, and what is the relevance of using it in front of other algorithms. Are other algorithms useful for doing such a detection job? E.g. the edge detection algorithm?
On page 7 false alarms procedures are not properly explained. If your images are static (same FOV) why not masking the areas occupied by buildings or other features producing false positives?
In conclusions you state that your algorith is able to detect about a 92% of the events, but what about the other 8%? Do you propose alternatives or other iterative procedures to increase such a percentage?
Minor points
Section 2, 1rst paragraph starts with a mistake: "Materials Generally, the grayscale, shape, size and shadow of aircraft objects...". The paragraph also includes other English incorrections.
English should be revised at some other points of the manuscript.
Additional references
Grav T. et al. (2011) PUBLICATIONS OF THE ASTRONOMICAL SOCIETY OF THE PACIFIC, 123:423–447
Raghunandan et al. (2018) IEEE - 7th International Conference on Communication and Signal Processing, DOI: 10.1109/ICCSP.2018.8524461
Soula S. et al. (2014) Atmospheric Research, Volume 135, p. 415-431.
Trigo-Rodríguez J.M., et al. (2008) Earth, Moon & Planets 102, 231-240.
Author Response
Date: August 3, 2020
Dear sir/madam,
Thank you for your review and comments on our paper applsci-890551. We have revised the manuscript according to your kind advice and detailed suggestions, and have highlighted the revised part. The point-by-point answers to the suggestions are listed as below.
Thank you very much for all your help, and we are looking forward to hearing from you soon.
Comments and Suggestions for Authors:
Point 1: The authors are presenting a paper discussing object detection in the field of computer vision, and particularly applied to the detection of aircrafts. The manuscript deals with object detection algorithms and its use in different fields (as stated in the introduction). The case is that the paper is very concentrated in aircraft detection, but ignores the use of these techniques in other research areas that are not properly cited. At least in the introduction should be explained.
Response 1:Thank you for your suggestion.
We have added the use of object detection techniques in other research areas like self-driving, biomedical image detection, automated defect inspection, satellite image analysis, criminal analysis.
Related information is added on lines 32-33 and 279-292 of the revised manuscript.
Point 2: Just to put an example, the authors are not citing any work related with the detection of asteroids, meteors and other atmospheric events. These fields have used computer vision techniques and detection algorithms as well and deserve to be cited (Trigo-Rodríguez et al., 2008; Grav et al., 2011; Soula et al. 2014).
Response 2:Thank you for your suggestion.
We have cited some work related with the detection of asteroids, meteors and other atmospheric events.
Related information is added on lines 30-31 and 272-278 of the revised manuscript.
Point 3: The introduction fails in explaining to a general reader what is an object detection algorithm. For example, a diagram could be useful (Raghunandan et al., 2018). In general the paper is not providing a case example either.
Response 3:Thank you for your suggestion.
Based on the literature suggested by the reviewer, we have explained what an object detection algorithm is. And we take an image in the Pascal VOC dataset [10] as an example to explain the meaning of object detection.
Related information is added on lines 33-36, 40-42, and 293-297 of the revised manuscript.
Point 4: They also fail to demonstrate how significant is the progress made by their initiative in comparison with other algorithms or techniques. On the other hand, it is unclear the exact contribution and novelty of this research to the field. You should put your results in an overall context. I also miss a general paragraph at the end of the introduction with the main goal of the paper.
Response 4: Thank you for your question.
In order to compare the test results with Faster R CNN, our dataset is annotated according to the label form of the PASCAL VOC2007 dataset. We train Faster R CNN on our training set and test on our test set. It can be seen from Figure 6 and Table 2 that the proposed method is better than Faster R CNN on our dateset.
We also revised the paragraph at the end of the introduction.
Related information is added on lines 227-230 and 82-87 of the revised manuscript.
Point 5: Section 2.2.1 Try to explain what is your contribution to the mean shift algorithm, and what is the relevance of using it in front of other algorithms. Are other algorithms useful for doing such a detection job? E.g. the edge detection algorithm?
Response 5: Thank you for your question.
Using the proposed aircraft object detection algorithm alone will detect multiple aircraft centers near the true aircraft center on a aircraft. The mean shift algorithm is used to cluster multiple aircraft centers into one. The edge detection algorithm alone is not good for weak and small aircraft objects. In the following research, we may consider combining the proposed algorithm with the edge detection algorithm and improving it.
Point 6: On page 7 false alarms procedures are not properly explained. If your images are static (same FOV) why not masking the areas occupied by buildings or other features producing false positives?
Response 6: Thank you for your question.
Our images are static (same FOV). Masking the areas occupied by buildings or other features producing false positives will definitely reduce false alarms. In the next research, we will study the algorithm for extracting the region of interest to further improve the accuracy of the algorithm.
Point 7: In conclusions you state that your algorithm is able to detect about a 92% of the events, but what about the other 8%? Do you propose alternatives or other iterative procedures to increase such a percentage?
Response 7: Thank you for your question.
Our algorithm is able to detect about a 92% of the events, and the other 8% can’t be detected mainly because the center of the circle on the aircraft don’t have the characteristics of four peaks and four valleys. We will propose a new algorithm to detect the remaining aircraft in our following research.
Related information is added on lines 260-263 of the revised manuscript.
Point 8: Section 2, 1rst paragraph starts with a mistake: "Materials Generally, the grayscale, shape, size and shadow of aircraft objects...". The paragraph also includes other English incorrections.
Response 8: Thank you for your suggestion.
Section 2, 1rst paragraph has been revised in lines 89-90.
Point 9: English should be revised at some other points of the manuscript.
Response 9: Thank you for your suggestion.
The article is reorganized further and English has been improved. Lines 30-36, 52-75 etc. have been revised.
Additional references
Grav T. et al. (2011) PUBLICATIONS OF THE ASTRONOMICAL SOCIETY OF THE PACIFIC, 123:423–447
Raghunandan et al. (2018) IEEE - 7th International Conference on Communication and Signal Processing, DOI: 10.1109/ICCSP.2018.8524461
Soula S. et al. (2014) Atmospheric Research, Volume 135, p. 415-431.
Trigo-Rodríguez J.M., et al. (2008) Earth, Moon & Planets 102, 231-240.
Author Response File: Author Response.pdf
Reviewer 3 Report
In the reviewed paper, the authors proposed an object detection algorithm. In general, I believe that this paper needs to be rewritten and proposed model described in more detail. Some of my concerns:
1. The introduction should be extended to show the problem and its possible application in practical problems. Moreover, machine learning solutions should be analyzed in more words then citing them (see line 44-52 on page 2).
2. Do not leave a section without any words (see 2.2)
3. In line 152, there is a mistake with a comma.
4. What means omega in Eq. (7).
5. Why is there an indent in someplace like after equations?
6. There must be more justification for your model. Add some visualization of data flow and example. Why is your proposal better than rCNN?
7. What about a situation with two objects on one image?
8. Some theoretical analyses (and profs) for time/computational complexities must be added.
9. The experimental section is without any sense. Adding some basic, commonly known equations is useless. Please, make a proper experimental test to prove that your proposal is good.
10. There is no comparison with other solutions. In the introduction you mentioned solutions like faster r-CNN, learning transfer, etc. Where is the comparison with them?
11. Where is a proper statistical analysis?
12. There must be made more tests on other datasets.
Based on the obtained results, I must ask the authors for showing the comparison with all the mentioned solutions from machine learning and object detection algorithms from the last 3-4 years. It is crucial for this paper.
Author Response
Date: August 3, 2020
Dear sir/madam,
Thank you for your review and comments on our paper applsci-890551. We have revised the manuscript according to your kind advice and detailed suggestions, and have highlighted the revised part. The point-by-point answers to the suggestions are listed as below.
Thank you very much for all your help, and we are looking forward to hearing from you soon.
Comments and Suggestions for Authors:
In the reviewed paper, the authors proposed an object detection algorithm. In general, I believe that this paper needs to be rewritten and proposed model described in more detail. Some of my concerns:
Point 1: The introduction should be extended to show the problem and its possible application in practical problems. Moreover, machine learning solutions should be analyzed in more words then citing them (see line 44-52 on page 2).
Response 1: Thank you for your suggestion.
In the case of few samples, machine learning methods will no longer be suitable for aircraft object detection. In this case, the proposed method can still accurately detect aircraft objects. This method can be used for aircraft object detection in remote sensing images. As we all know, remote sensing images are difficult to obtain and very expensive. Therefore, it is very cost-effective to use the proposed method to detect aircraft objects in remote sensing images. In civilian use, the proposed method can be used for airport management, for example, aircraft scheduling, aircraft statistics, etc. In military use, it can be used to detect aircraft objects and improve early warning capabilities.
We have re-analyzed machine learning solutions and given a brief introduction to each solution.
Related information is added on lines 52-75 of the revised manuscript.
Point 2: Do not leave a section without any words (see 2.2)
Response 2: Thank you for your suggestion.
We have added a paragraph in Section 2.2
Related information is added on lines 174-177 of the revised manuscript.
Point 3: In line 152, there is a mistake with a comma.
Response 3:Thank you for your suggestion.
We have corrected this error. Related information is added on line 186 of the revised manuscript.
Point 4: What means omega in Eq. (7).
Response 4: Thank you for your suggestion.
is the weight of the sample , and .
Related information is added on line 198 of the revised manuscript.
Point 5: Why is there an indent in someplace like after equations?
Response 5: Thank you for your question.
We have removed all indent after all equations in the revised manuscript.
Point 6: There must be more justification for your model. Add some visualization of data flow and example. Why is your proposal better than rCNN?
Response 6: Thank you for your suggestion.
We have added some example to justify our model. Based on pre-trained backbone network VGG16, Faster R CNN is trained on our training set. It can be seen from the visualization results and some evaluation indicators that the proposed method is better than Faster R CNN in the case of small sample.
Related information is added on lines 134-135, 227-230 and 364-365 of the revised manuscript.
Point 7: What about a situation with two objects on one image?
Response 7:Thank you for your suggestion.
If there are two objects of different sizes in the image. We use the proposed algorithm first to detect the large and saliency aircraft. Then the part of these aircraft on the original image can be deducted. Finally, a small radius is chosen and small aircraft can be detected by our method.
Point 8: Some theoretical analyses (and profs) for time/computational complexities must be added.
Response 8: Thank you for your suggestion.
We calculated the average time it takes for the proposed algorithm and Faster R CNN to detect images on the test set. It can be found that our method is faster than Faster R CNN.
Related information is added on lines 251-252 of the revised manuscript.
Point 9: The experimental section is without any sense. Adding some basic, commonly known equations is useless. Please, make a proper experimental test to prove that your proposal is good.
Response 9: Thank you for your suggestion.
We have removed these equations and added a proper analysis.
Related information is added on lines 248-251 of the revised manuscript.
Point 10: There is no comparison with other solutions. In the introduction you mentioned solutions like faster r-CNN, learning transfer, etc. Where is the comparison with them?
Response 10: Thank you for your suggestion.
In order to compare the test results with Faster R CNN, our dataset is annotated according to the label form of the PASCAL VOC2007 dataset. We train Faster R CNN on our training set and test on our test set. It can be seen from Figure 6 and Table 2 that the proposed method is better than Faster R CNN on our dateset.
Related information is added on lines 227-230 and 232-235 of the revised manuscript.
Point 11: Where is a proper statistical analysis?
Response 11: Thank you for your suggestion.
We have revised the statistical analysis of our paper.
Related information is added on lines 236-253 of the revised manuscript.
Point 12: There must be made more tests on other datasets.
Response 12: Thank you for your suggestion.
We have added the plane dataset in UCAS_AOD dataset to our dataset, which includes 1000 images and a total of 7482 airplanes.
Related information is added on lines 125-135 of the revised manuscript.
Point 13: Based on the obtained results, I must ask the authors for showing the comparison with all the mentioned solutions from machine learning and object detection algorithms from the last 3-4 years. It is crucial for this paper.
Response 13: Thank you for your suggestion.
Due to time constraints, we chose the most representative algorithm Faster R-CNN in recent years to compare with the proposed algorithm. It can be found that in the case of small samples, the proposed method has much better detection performance than Faster R-CNN, and the detection speed is faster. It can be seen from the data analysis that the proposed method can even detect aircraft objects without samples. Therefore, in the case of small samples, the proposed method is better than Faster R-CNN.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The paper has been improved on the basis of the given suggestions.
Reviewer 2 Report
Dear authors
I think that your answers are enough to satisfy the reader curiosity about some of your main claims. You also improved the introduction.
I only noticed some minor issues in the reference list. Please be sure that you cite properly the authors: surname, N.
For example reference 2 is wrongly cited, it should be: Trigo-Rodriguez J.M. et al.
Please double check the references before the publication of the paper
Thanks
Reviewer 3 Report
Accept
