FESSD: Feature Enhancement Single Shot MultiBox Detector Algorithm for Remote Sensing Image Target Detection

Round 1

Reviewer 1 Report

Although The article is well organized and written well. However I have noticed few things.

1. Abstract does not include any kind of results achieved in terms of recall precision and other parameters, if possible, it should be included

2.The Grammatical errors should be removed

I Recommend it for the publication in this journal

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This is an interesting paper for target detection of remote sensing images based on deep learning.

The following are some comments and suggestion for possible publication.

1. In abstract, the authors should describe the performance (mAP %) of proposed methods on the two datasets((SD-RSI,DIOR).

2. In section 3.1. Overall Architecture, the authors should write the full form of SSD before using the acronym. [Single Shot MultiBox Detector(SSD)].

3. In 4.5. Performance Evaluation and Comparison, please compare the performance of proposed method with "FFESSD: An Accurate and Efficient Single-Shot Detector for Target Detection" by Wenxu Shi,et al.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The article presents a new framework for target detection in remote sensing images based on the latest achievements in single shot detectors. The authors connected the state-of-the-art frameworks together in one architecture which enhances the single shot multi-box detectors as an algorithm, accomplishing very good results. However, the novelty of the paper is slightly limited because the authors used conventional elements, like puzzles.

The strong points significantly exceed the weak ones, so my evaluation is very high and I strongly recommend publishing this article.

Strong points:

Not only are the results obtained very good, but also the time of calculation and the number of parameters are the lowest compared with the methods shown in tab. 6.

The authors uncommonly compared all important elements such as:

• visual detection results of different feature enhancement modules,

• the feature maps visualization results of the last convolution layer,

• the effect of different attention mechanisms on the model training process and detection accuracy,

• the heat maps of different attention mechanisms used in the research

which are rarely analysed and published.

Weak points:

The number of detected object classes is limited.

The programming language/modules have not been identified.

Detailed remarks or/and publishing comments :

1. l. 65: self-attention mechanism → reference is missing

2. Structures in fig. 2 and fig. 11 are too little and illegible. It would be better to show each structure in a separate figure and make them bigger.

3. Fig. 11 and 12 have the same caption, which is irrelevant.

4. English should be proofread.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf