A Novel Ensemble Based Reduced Overfitting Model with Convolutional Neural Network for Traffic Sign Recognition System
and Manuel Mazzara 2Round 1
Reviewer 1 Report
Good work. Please solve the following comments with a major revision and to be reviewed again.
1. Some full names of abbreviations are necessary. Some examples:–Page 1 line 12: “ELVD”–Page 1 line 17: “GTSD”. I will not list all here but the paper needs to go through another round of proofreading.
2. In Abstract, the authors claimed that the ELVD model is better than an average human. Add relevant explanation. In the introduction, when discussing the related works on traffic monitoring fields in Introduction, the authors should provide richer works, such as sar shadow tracking, sar ship instance segmentation and sar ship detection: shadow-background-noise 3d spatial decomposition using sparse low-rank gaussian properties for video-sar moving target shadow enhancement, a mask attention interaction and scale enhancement network for sar ship instance segmentation, a group-wise feature enhancement-and-fusion network with dual-polarization feature enrichment for sar ship detection, depthwise separable convolution neural network for high-speed sar ship detection, high-speed ship detection in sar images based on a grid convolutional neural network, and so on.
3. The statement of contributions at the end of Sec. I should be enriched more, also aiming at better highlighting the relevant technical challenges tackled by the authors.
4. Increase the readability of Fig. 5.
5. It is difficult to tell the difference between Fig. 9 (c) and Fig. 9 (d). Maybe it would be better to change another traffic sign figure.
6. Add the explanation of “Support” index used in Table 5.
7. Add more comparison experiments with other SOTA methods.
8. Explain why the proposed ELVD can achieve fastest average detection time.
9. IMHO, the Conclusion should be re-written to 1) explicitly describe the essential features/advantages of the proposed method that other methods do not have, 2) describe the limitation(s) of proposed method, and 3) what aspect(s) of the proposed method could be further improved, why and how.
10. The English should be improved greatly.
Author Response
The corrections suggested by the reviewers has been incorporated in the updated version of the paper.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
I read the paper carefully. Unfortunately, I can’t accept the paper it its current form. The authors should address the following comments with highlighting the importance of the study and their contributions in the field in terms of new idea in the models and methods and not just using.
1. The title should be changed to be clear and concise. Try to avoid using acronyms if possible.
2. There is no Introduction section. The authors can create this section and explain the importance of the research and highlight the main contributions of the study.
3. Another new section should be created entitled "literature review" or "related works".
4. In literature review section, the authors should include the recent related methods used for traffic sign recognition and should provide a clear discussion on what is the gap in these work on the same problem and how the authors fill this gap by this research work.
5. There is no comparison with the other published related work to show the improvement presented by this work. Statistical hypothesis is required to show whether the improvement is significant or not.
6. There is no comparison with some other deep learning models and architectures.
7. An optimization method or algorithm is required to select the best values of hyper-parameters rather than trying various values.
8. How to justify that the model is robust to noisy input data.
9. Please summarize in a Table the number of instances used for training, validation, and testing.
10. Extensive editing of English language and style are required. The manuscript has many writing errors.
11. Ensemble methods greatly increase computational cost and complexity and traffic sign recognition should be suitable for real-time recognition. I think this issue make the proposed solution is not suitable than lightweight model.
Author Response
The corrections suggested by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
Reviewer 3 Report
1. The proposed method is an ensemble of well-known existing methods, LeNet5, VGGNet, and Dropout Net. The referee recommends that the authors explain more clearly how their ensemble method is formed instead of explaining LeNet5, VGGNet, and Dropout Net. In this sense, The referee would like to recommend removing the parts that explain LeNet5, VGGNet, and Dropout Net from the manuscript.
2. The referee needs to learn what it stands for, 'VGGANET' in the title. The referee guesses that 'VGGANET' is a typo of 'VGGNET.'
3. Figure 2 is a scanned image from LeCun's paper. This type of citation requires a specified courtesy.
4. In equations 1) and 2), it is better to specify the definition of the loss function.
5. In eq. 1), L2 regularization term is used, but it is called L1 regularization. Similarly, in eq. 2, L1 regularization term is used, but it is called L2 regularization.
Author Response
The corrections suggested by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Some comments are not solved well, e.g., the authors declare that added in some places, but this reviewer cannot find any changes, so a major decision is still needed.
Author Response
The corrections suggested by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
Reviewer 2 Report
Unfortunately, I can’t accept the paper it its current form. The authors didn't address the major comments.
Author Response
The corrections suggested by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors revised the thesis by correcting the contents pointed out by the reviewer. The revision improved the presentation of the manuscript.
Author Response
The corrections suggested by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
Round 3
Reviewer 1 Report
Some comments are "STILL" not solved well, e.g., the authors declare that added in some places, but this reviewer cannot find any changes, so a MAJOR decision is still needed.
Author Response
The comments mentioned by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
Reviewer 2 Report
The novelty of the work is limited. The authors just combined (Ensemble) what already existed models to recognize traffic sign images, which make the system more complex and computationally inefficient. Such system should be real-time system. The efficiency is very important for traffic sign recognition system. There are many deep learning models have been published in several papers for traffic sign recognition but the authors didn’t compare their work with them and highlight their contributions.
You can search and find many papers on the same datasets, for example, you can see one paper published in 2011, achieved a high accuracy rate on GTSRB dataset
Sermanet, Pierre, and Yann LeCun. "Traffic sign recognition with multi-scale convolutional networks." In The 2011 international joint conference on neural networks, pp. 2809-2813. IEEE, 2011.
For the previous comments, still, the authors can’t address the major comments 5, 7, and 8, listed below:
5. There is no comparison with the other published related work to show the improvement presented by this work. Statistical hypothesis is required to show whether the improvement is significant or not. (Where is the comparison and where is the statistical hypothesis)
7. An optimization method or algorithm is required to select the best values of hyper-parameters rather than trying various values.
8. How to justify that the model is robust to noisy input data. (The answer should contain a case study to justify).
Moreover, the paper contains some writing errors. For example in Line “blurr” 656
Also, the authors repeated what already existed in Table 7 in Figures 12.a and 12.b. They said with “existing methods”?! In any work these methods are existing. In comment 5, I asked the authors to compare with existing methods in some recent related work.
Author Response
The comments mentioned by the reviewers has been incorporated in the updated version of the paper.
Author Response File: Author Response.pdf
