A Marine Small-Targets Classification Algorithm Based on Improved Convolutional Neural Networks
Round 1
Reviewer 1 Report
In this paper, convolutional neural network (CNN) techniques were combined with the IRC module to improve the ship classification accuracy and training efficiency. Based upon the results, the average accuracy of ship classification of the model reaches 98.71%, which is about 3% more accurate than the traditional network model and about 1% more accurate compared with some recent improved models. The results can be applied to marine ship classification management work in the future. I have following comments on the paper before the paper can be accepted.
(1) English used in the paper should be significantly improved. For example, the title of the paper doesn’t make sense to me. What is complex working conditions? There is no description. From Lines 77 to 82, the sentences is really too long, and two “in order to ” were used. Similarly, Line2 259 to 264. Please shorten the sentences in the whole paper and check some grammar mistakes.
(2) The novelty of the paper may be “IRC module”. What does IRC stand for? And it is very weird, np mention of the “IRC module” in the abstract. If it is the main novelty, why does not mention the “IRC module” in the abstract and key words.
English used in the paper should be significantly improved.
Author Response
Please see the attachment
Author Response File: 
Author Response.doc
Reviewer 2 Report
Dear Authors
This paper proposes a Research on Complex Working Condition Small Target Recog-2 nition Algorithm Based on Deep Learning. The paper is ok, but in my opinion it lacks a better description of the state of the art and a better explanation of the novelty of this work. Especially, the authors should compare their precision and resolution results with the literature in order to demonstrate their contribution
Page 1 define IRC, and all tha acronyms in the manuscript.
page 1, lines 27-29. With the development of China's economy and technology, more and more attention is paid to maritime safety, and more attention is paid to the management classification of maritime vessels, and the study of small target identification for safety monitoring of hydraulic structures such as ports and dams. according to whom? include a reference
page 4, how do the authors achieve the equation 1-9? Please give more details or a reference.
Page 3, line 100: why the authors are introducing the elements in equation 2, 4, 5,6 ? define all the elements.
Page 5, lines 192-197: I cannot understand the explanation. Please better explain it, or give a reference
Page 7, lines 252-253, The model accuracy can reach up to 98.79% with a fluctuation range of 0.3%, and the model loss value is about 0.0351 with a fluctuation range of 0.005%. compared with which model.
page 7, lines 261-264. How much more time is necessary?
Please improve the conclusions?
Dear Author
improve this section
The IRC module structure differs from the Inception-ResNet [17] are in discarding the 1.1 convolutional branch and adding a maximum
pooling layer branch. The 1convolutional branch is discarded because this convolutional layer cannot form the origin identity, and the convolution branch is used to reduce the dimensionality of the feature map, but the IRC architecture requires the output dimension to be the same as the input dimension, so there is no need for the branch to change the input dimension
Author Response
Please see the attachment
Author Response File: Author Response.doc
Reviewer 3 Report
The proposed article is interesting, although there are not many new elements. There are several aspects that need further investigation.
1. It is important to better delineate the design differences of the proposed model from the known models with which it has been compared, as it is not clear what features allow the proposed model to perform better than others.
2. In order to better define the accuracy of the results obtained, in addition to the table showing precision, recall and F1 score, it would be interesting to include the confusion matrix, to get an idea of what false negatives and positives are highlighted.
3. The discussion of results and the conclusion certainly need to be expanded and more clearly written
4. It would be interesting to include tests of use of the model on real ports
5. The term KNN (page 2) was not previously defined
There are some long sentences, which could be rewritten more clearly
Author Response
Please see the attachment
Author Response File: Author Response.doc
Round 2
Reviewer 1 Report
I am satisfied with the revised manuscript.
I am satisfied with the revised manuscript.
Reviewer 2 Report
Dear Authors
I think the manuscript looks better
Reviewer 3 Report
The answers provided by the authors fully addressed the comments. The paper can be accepted in the present form
