A Fast Maritime Target Identification Algorithm for Offshore Ship Detection
Round 1
Reviewer 1 Report
Add more relevant plots for with mathematical justification
Also include latest reference for the state of art comparison which will further justify the novelty of the work
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper proposed a Fast Maritime Target Identification algorithm, FMTI, to identify maritime targets rapidly. The FMTI adopts a Single Feature Map Fusion architecture as its encoder, thereby improving its detection performance for varying scales of ship targets, from tiny-scale targets to large-scale targets. T
a) On the other hand, the passive approach is usually used for the ship's 37
au-tomated identifying system. ( suggest automated)
b) Table 2 Additional case results, table is split into two page . ( suggest to realign )
c) There are 4267 images totally in the dataset, with 20% designated 280
for the test set, and the rest for the training set. A bit too small to the machine learning and deep learning. (suggest to download public database and test it as benchmark to the preliminary results).
d) with 20% designated 280 for the test set, and the rest for the training set. ( suggest to try on other train test ratio such as 50-50, 60-40, 70-30)
e) Figure 4. Recognition results of FMTI. (A) multi-objective, (B) Simple or Single objective. and table 4 Preliminary comparison results of different combinations are out of alignment.
f) Table 4 was obtained in the 2017 COCO validation set, and table 5 was acquired from self-built datasets. ( Table 5)
g) lack of analysis why FMTI algorithm is better than others.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The manuscript entitled “A fast maritime target identification algorithm for offshore ship detection” proposes a Fast Maritime Target Identification (FMTI) algorithm based on the Single Feature Map Fusion architecture as its encoder. The experiments conducted on the custom dataset with images of maritime objects (ships) show good performance, which also surpasses that of the YOLOF method.
The proposed method is straightforward and oriented to practical application. The results are obtained by testing it on the corresponding dataset and compared to those obtained by the competitive detection method. The conclusions appropriately address the limitations of the presented study and also provide the directions for future research.
Here are some comments I would like the authors to address before the manuscript is considered for publication:
- The manuscript needs additional (and almost extensive) proofreading and correction of grammar and language. Due to the lower level of English, the manuscript is unfortunately very difficult to read and understand the authors’ presentation in some places. Additionally, throughout the manuscript, many words are unnecessarily separated by the hyphens in the middle of them.
- The authors correctly state the state-of-the-art performances of the CNNs in many different applications today. Please support these with some of last year’s studies to briefly illustrate and provide an interested reader with examples of very diverse applications. This can be done for the ResNet architecture, particularly as it was used as a backbone in the tested methods. Please consider mentioning the following papers: 10.3390/diagnostics11010105; 10.1109/ACCESS.2021.3139850; 10.3390/e23040435; 10.3390/diagnostics11050893.
- Please provide more detailed information on the training process and employed training parameters.
- Please also elaborate on the measures taken to prevent the overfitting of the models.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
I congratulate the authors on the work done.
In general, the literature review is sufficient. However, it would be useful to refer to several papers published in 2021 or 2022. This is important given the technical work carried out, which is developing dynamically. The contribution of the authors is clearly stated in the introduction.
The description of the methodology is also done sufficiently. Several pieces of information should be completed. First of all, explain the designations used in the models. Even if they are obvious (e.g. in Formula 6)
The description of the method of obtaining images for analysis should be developed in the paper.
Do the authors plan to provide a link to the input data? It would allow other researchers to verify their research work and make comparisons of results for the same input base.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The paper is revised and all the queries are completed. From my side it is ok to publish.
Reviewer 2 Report
from authors, "We attempted the 50-50 ratio, but the model did not perform as well as anticipated (mAP<32.9%). And we also tested the 60-40 ratio and the results were unsatisfactory (mAP<36.3%). ". if possible, still present the experiment results for others train-test ratio, to show that the large sample size of dataset is important and work well even on different ratios. as deep learning required large dataset.
Reviewer 3 Report
The authors have addressed my comments.
