LFLD-CLbased NET: A Curriculum-Learning-Based Deep Learning Network with Leap-Forward-Learning-Decay for Ship Detection
Round 1
Reviewer 1 Report
The paper deals with the problem of training a neural network model to implement a ship detection task.
The topic tackled by the authors is interesting and worth being considered due to its relevant impact in nowadays maritime policy and decision-making frameworks. Despite this, some critical points need to be addressed.
1) Usage of acronyms: either in the title as in the abstract acronyms should be avoided. Sometimes in the paper acronyms are also directly introduced, without explanation. It is not unusual to have multiple meanings for the same acronym. An explicit definition ensures to avoid such ambiguity.
2) Section 3.4: the acronym is changed from LFLD to LDLF.
3) For this kind of problems, the data typology is not a secondary matter. In the ‘Introduction’ Section the authors state that one of the obtained results is a “small real maritime ship detection dataset”, but it is never well described in the text. E.g. what kind of sensors provide the data you analyze? Is it radar? Optical? Both? Different sensors choices give birth to different issues to tackle, so it is important to include in your study a thorough discussion concerning the data typology you chose and the pros and cons implied by such a choice.
4) Once more about the dataset: in Section 4 the authors state that the dataset is freely available but they do not provide any information concerning how to retrieve such data. Please include the instructions to obtain the dataset you exploited. This is important in order to reproduce the experiments you have performed, in case that anyone interested wishes to confirm the promising performance that you claim in the ‘Experiments and Results’ Section.
5) The authors state that the model training is based on an incremental complexity criterion, but they do not clearly explain the concept of complexity and the way the related increase is implemented. Please provide detailed explanations.
6) In equation (6) the upper-case symbol W_t appears, but later a lower-case w_t is defined. Please avoid ambiguities, especially in the math descriptions.
7) Plots in Section ‘Experiments and Results’ are way too small. Please make them more readable.
8) In Section 4.10 the authors mention Table 4. There is no Table 4 in the paper.
The opinion of the reviewer is that the paper deserves further reworking before resubmitting to JMSE.
The paper's English quality is just enough sufficient. A considerable number of mistakes and low-quality phrasing are present throughout the text. An English language review is strongly suggested.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Dear Authors,
This work presents a novel ship detection model, LFLD-CLbased NET with incorporated curriculum learning and learning rate decay method. In order to raise the value of the article, please pay attention to these issues and make every effort to improve and supplement article.
- Why did you choose LFLD-CLbased NET? Did you conduct any preliminary scientific research which proved better results of ship detection? If you did it, please explain in the text of manuscript.
- It is unclear what method you chose to set percentage of training data set and test data set in tables 1 and 2. If we better look in Table 1, the percentage of Training Deep Learning vessel dataset is about 96%, while in Table 2 the Test Real-world Ship Detection Dataset is about 50%. Please, explain it in text of manuscript.
Kind regards
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I received the cover letter and appreciated the reported modifications. From my side, I confirm that the changes made fulfill my requests and suggestions and that the novel version of the paper is now suitable for publication on JMSE.
English quality is sufficient now
