Absorption Pruning of Deep Neural Network for Object Detection in Remote Sensing Imagery
Round 1
Reviewer 1 Report
An absorption pruning method for solving overparameterization problem in DCNN is proposed via filter selection and adding absorb loss, and the CenterNet101 is taken as a backbone network. The research topic is interesting, and the experiments are suitably designed. However, the writing quality is far from the standard of publication. There are many reductant and unsuitable expressions. It is strongly suggested the authors to proof reading the manuscript carefully. Moreover, the proposed method is difficult to follow. Following are some detailed questions and suggestions from the reviewer:
1. Please explain what the “filter parameter tensor” is and how is the converged parameter tensor P_i,j^* obtained? In filter selection step, the model is not converged. Is it from well-trained unpruned model? Authors should make it clear.
2. Figure 2: What are the 10 filters？Stochastic-initialized or something else? Their concrete form should be explained.
3. The theoretical basis of designing a linear function and adding its L1-norm parameters vector as loss_ab (Equation 4 and 5) should be introduced.
4. Figure 6 and 7: "le-3" may be "1e-3"? And the curve of Loss and Loss* (Equation 2 and 3) are advised to be added and compared.
5. Figure 9: What do the red and blue labels represent? And the accuracy is hardly visible in the top 2 images. Some labels are overlapped.
6. In Figure 9, authors only displayed the result of pruned CenterNet101. They are advised to compare and display the detection result images of both original and pruned network to prove the accuracy-preserving capability of the proposed method in the light weighted network. And more results should be displayed.
7. What is gama and beta i in the equation (5)
8. Please explain the rationale of adjusting the pruning ratio by equation (6) and (7), in other words, why they have the present form.
9. Some details in the pseudo code is missing, such as step 3.
10. Line 55: A space key is missing.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Summary:This is an interesting work.
The recognition of objects in SAR images is a hot topic in this direction in recent years. More and more neural network target recognition methods originally used in optical images are applied to SAR images. However, the recognition methods applied to traditional optical images are structurally redundant for SAR images.
In order to obtain lightweight remote sensing object detection neural networks, this manuscript proposes a network pruning method that takes into account the characteristics of remote sensing objects, named absorption pruning. Unlike the classical iterative pruning pipeline used by existing pruning methods, this manuscript innovatively proposes a four-step pruning pipeline that only needs to be executed once. At the same time, the criteria for selecting absorption filters a pruning ratio adjustment method based on object characteristics in remote sensing images to optimise the design of the pruning method are given in this manuscript.
The present form of the manuscript is well written, although some minor problems need to be addressed.
The paper can be improved in the following aspects:
Detailed comments:
1. In section 1, you said“Finally , fine-tuning will be used to recover the network performance. Ultimately, the absorption pruning method will output a light object detection network.” The “Finally” and “Ultimately” two words have the same meaning, which will cause some reading problems when put together.
2. In section 2, you said“Network pruning has been proven effective in reducing network complexity while preserving network performance.” Please list some relevant articles.
3. In section 3, I think you can modify Figure 2 to explain why the number of epoch equals 10.
4. I suggest explaining why the total training epochs are set to 60 in section 4.1.3.
5. I suggest to use relevant data in the conclusion to prove the effectiveness of the proposed method more intuitively.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Extensive Englished is still needed for this paper. For instance:
Line 407: "takes second place", should be: "takes the second place"
Line 307:" Right: SAR image with ship object." should be: "Right: SAR image with ship objects."
There are also some sentences that are too long to understand.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
I’m very glad to receive your revised manuscript.
In this revised version, the questions raised after the last review have been modified. In general, the problems raised last time have been properly modified and marked in the original text, but there are still some minor problems. So, I will make a brief comment on your revision.
You have made satisfactory modifications to the questions 1, 2 and 5, making the article easier to understand, and making readers more intuitively understand your work and the advantages of the proposed model. For question 3, although you has been added with a new annotation, which roughly shows the position where the number of epoch equals 10, but it is not very intuitive. Your explanation of question 4 in the reply solves my doubts, but the corresponding parts of the original text have not been modified. I wonder if you can add a small amount of explanation to the manuscript.
Your article has no major problems in general, but some details can be better.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
