Comparison of Multilayer Neural Network Models in Terms of Success of Classifications Based on EmguCV, ML.NET and Tensorflow.Net
Round 1
Reviewer 1 Report
The study is generally successful. The mathematical background of the study can be developed. In addition, the purpose of the study should be clearly emphasized in the summary and conclusion sections. I would like you to add the following 2 references that can be developed to the Reference section.
1- https://link.springer.com/article/10.1007/s00521-021-06652-w
2- https://link.springer.com/article/10.1007/s00521-020-05436-y
Author Response
Dear Reviewer, Thanks for the revisions and helpful suggestions. Based on your review, we tried to improve the paper by addressing the following points:
The use of neural networks is now standard. We have to agree with the reviewer on this. Since convolutional neural networks are used by default in the classification phase, it makes no sense to state the mathematical foundation (in the past, yes, today it would only be a copy of statements presented several times). What makes the paper new is the use of the EmguCV library. We have added the following explanation to the conclusion section:
In this paper, we used convolutional neural networks in individual models (EmguCV, ML.NET and Tensorflow.Net). The use of convolutional networks makes it possible to train networks with high classification success in various application areas, for example in Detection of invisible cracks in ceramic materials [47] or Classification of operation cases in electric arc welding wachine [48].
EmguCV library is an extension of the OpenCV. From the available databases of professional articles (e.g. Scopus), we can see that the issue of applying EmguCV is more oriented to its use in the first phase of the recognition process – detection [49], [50], [51]. Currently not exist many relevant sources that refer to extraction or classification phases. In this paper, we compared different ways of classification EmguCV, ML.NET and Tensorflow.Net. The last 2 mentioned are among the most used. However, as we state in the paper, EmguCV contains cascade classificators very similar to the one used in the application of the Viola-Jones algorithm. In the article, we therefore showed that we can successfully use the EmguCV model to train and test the speed also in the classification phase.
Reviewer 2 Report
The paper proposes a comparison of some models of multilayer neural networks in the person's face classification.
I cannot detect elements of novelty in the manuscript.
The section on neural networks is not described formally and mathematically and contains various imprecise sentences. More details on the various layers used should be provided.
The results section is poor and should be enriched.
I suggest rejecting the manuscript.
Author Response
Dear Reviewer, Thanks for the revisions and helpful suggestions. Based on your review, we tried to improve the paper by addressing the following points:
The use of neural networks is now standard. We have to agree with the reviewer on this. Since convolutional neural networks are used by default in the classification phase, it makes no sense to state the mathematical foundation (in the past, yes, today it would only be a copy of statements presented several times). What makes the paper new is the use of the EmguCV library. We have added the following explanation to the conclusion section:
In this paper, we used convolutional neural networks in individual models (EmguCV, ML.NET and Tensorflow.Net). The use of convolutional networks makes it possible to train networks with high classification success in various application areas, for example in Detection of invisible cracks in ceramic materials [47] or Classification of operation cases in electric arc welding wachine [48].
EmguCV library is an extension of the OpenCV. From the available databases of professional articles (e.g. Scopus), we can see that the issue of applying EmguCV is more oriented to its use in the first phase of the recognition process – detection [49], [50], [51]. Currently not exist many relevant sources that refer to extraction or classification phases. In this paper, we compared different ways of classification EmguCV, ML.NET and Tensorflow.Net. The last 2 mentioned are among the most used. However, as we state in the paper, EmguCV contains cascade classificators very similar to the one used in the application of the Viola-Jones algorithm. In the article, we therefore showed that we can successfully use the EmguCV model to train and test the speed also in the classification phase.
Reviewer 3 Report
It is really necessary to have such old references to explain neural networks like 1943 and 1958 years.It is very easy to lose the thread of the text due to the way of citing some of the references, perhaps it is better to include the name of the first author and the number of the reference.
Author Response
Dear Reviewer, Thanks for the revisions and helpful suggestions. Based on your review, we tried to improve the paper by addressing the following points:
Thanks for the warning. The main aim of introducing the years 1943 and 1958, we wanted to draw attention to the fact that neural networks alone do not represent a new problem. However, the possibilities and ways of their use are very wide and since 1943 they have been used successfully in various fields. Of course, in the References section, these years are not used, but are interpreted by the latest sources.
Unfortunately, we cannot influence the method of citation, this is how the template is set and we as authors must adhere to it.
Reviewer 4 Report
This draft could probably be used as a survey paper, as the methods in discussion are standard without too much new contribution. Also, the database is small and the images are mostly frontal images, so the accuracy result may not be representative. One could use the FERET database, to make the results more convincing.
Author Response
Dear Reviewer, Thanks for the revisions and helpful suggestions. Based on your review, we tried to improve the paper by addressing the following points:
The use of neural networks is now standard. We have to agree with the reviewer on this. Since convolutional neural networks are used by default in the classification phase, it makes no sense to state the mathematical foundation (in the past, yes, today it would only be a copy of statements presented several times). What makes the paper new is the use of the EmguCV library. We have added the following explanation to the conclusion section:
In this paper, we used convolutional neural networks in individual models (EmguCV, ML.NET and Tensorflow.Net). The use of convolutional networks makes it possible to train networks with high classification success in various application areas, for example in Detection of invisible cracks in ceramic materials [47] or Classification of operation cases in electric arc welding wachine [48].
EmguCV library is an extension of the OpenCV. From the available databases of professional articles (e.g. Scopus), we can see that the issue of applying EmguCV is more oriented to its use in the first phase of the recognition process – detection [49], [50], [51]. Currently not exist many relevant sources that refer to extraction or classification phases. In this paper, we compared different ways of classification EmguCV, ML.NET and Tensorflow.Net. The last 2 mentioned are among the most used. However, as we state in the paper, EmguCV contains cascade classificators very similar to the one used in the application of the Viola-Jones algorithm. In the article, we therefore showed that we can successfully use the EmguCV model to train and test the speed also in the classification phase.
We agree to the use of the FERET database. We also used this database during the experiment (which we do not mention in the paper). However finally, we chose the YALE database. This is because the FERET database contains a number of identical (repetitive) images and the big problem is the network is re-learned. We consider this to be a very serious problem, which we want to report on in the future.
Round 2
Reviewer 4 Report
The revision has been minimal and this draft is not too significant.
