An Efficient Smoke Detection Algorithm Based on Deep Belief Network Classifier Using Energy and Intensity Features

Round 1

Reviewer 1 Report

The study attempted to detect the smoke in video using the Deep Believe Network classifier and compare the results with existing algorithms. The manuscript is prepared nicely, though few language issues are to address.

Though the proposed methodology has a high detected rate of 96%, the experimental results as shown in table 5 CNN slightly outperformed the proposed method in both with and without noise cases. Indeed, the performance of all three methods compared, SVM, DBN, and Deep CNN is already above 90% therefore, improving 1 or 2% may not be the real improvement, it may be due to the context, noise, or other parameters in an input dataset. The major contribution of this work could be computation time and resources required. Authors mentioned in line 246 – 249 that the proposed method is simpler than the existing one, what does mean simpler? I would like to see a performance matrix (not only the accuracy) of all the methods compared in this manuscript before deciding, only accuracy is not enough to make a decision in this context. Apart from this, the following minor issues to address in the manuscript.

 

Typo in the caption of figure 11

Line 552: there is some grammatical error.

Line 597: what does it mean? Doses it means, this method is not the best performer, right? If it not superior than the existing method(s) in any aspect, then what is the point to publish this paper? Only to show an alternative, which is inferior to existing, is not enough reason to publish as a research article.

Author Response

Answers to Reviewers 1 comments
Manuscript title: An Efficient Smoke Detection Algorithm Based on Deep Belief Network
Classifier Using Energy and Intensity Features
Authors : Rabeb Kaabi , Moez Bouchouicha, Aymen Mouelhi , Mounir Sayadi, Eric
Moreau
We are very grateful to the Reviewer 1 for his several corrections and recommendations
which greatly improve our manuscript. We agree with all the reviewer’s comments and have
corrected the suggested errors and mistakes and we have done the appropriate changes to
fulfil the comments. Below, we provide specific and detailed responses for each question.
Question 1: Though the proposed methodology has a high detected rate of 96%, the
experimental results as shown in table 5 CNN slightly outperformed the proposed method in
both with and without noise cases. Indeed, the performance of all three methods compared,
SVM, DBN, and Deep CNN is already above 90% therefore, improving 1 or 2% may not be
the real improvement, it may be due to the context, noise, or other parameters in an input
dataset. The major contribution of this work could be computation time and resources
required. Authors mentioned in line 246 – 249 that the proposed method is simpler than the
existing one, what does mean simpler? I would like to see a performance matrix (not only the
accuracy) of all the methods compared in this manuscript before deciding, only accuracy is
not enough to make a decision in this context. Apart from this, the following minor issues to
address in the manuscript.
Response 1: Line 672-679
The advantages of the proposed method with respect to other smoke detection methods can be
summarized in two points. The first one is the easy and fast localization of the smoke regions
with an IoU changing from [0.85 to 0.94] for videos containing smoke regions. Smoke
detection using the proposed method can easily and rapidly localize and classify smoke and
no-smoke regions compared to the conventional smoke detection methods and Deep CNN
that requires a lot of computation time as well as a lot of data, which is sometimes not easily
available. The calculated time of training process is presented in the Table 5.
The method that we propose is simpler in terms of algorithm since Deep CNN injects 2 CNN
networks and a classifier SVM to classify the smoke regions. Therefore, our method
consumes fewer resources and is characterized by the highest computation time.
The real value of our work is to locate and classify regions of interest with interesting values
of performance and the computation time and resources needed for training and testing
without forgetting that the main purpose of smoke extraction is to locate and classify with a
minimum of execution time and resources and high performance metrics.
The performance matrix and any necessary changes have been made and are presented in
Tables 3 and 5.
Question 2 : Typo in the caption of figure 11
Answer 2: Done
Question 3 : Line 552: there is some grammatical error.
Answer 3: Done
Question 4: Line 597: what does it mean? Doses it means, this method are not the best
performer, right? If it not superior to the existing method(s) in any aspect, then what is the
point to publish this paper? Only to show an alternative, which is inferior to existing, is not
enough reason to publish as a research article.
Answer 4: CNN-based smoke detection methods are significantly slower due to an operation
such as maxpool and the training process takes a lot of time if the computer does not consist
of a good GPU. So, the contribution of this work lies in the fact that we worked on the
preprocessing step on new combination of features: smoke-color, smoke-motion, and energy
to extract the regions of interest that we inserted within a simple architecture. Our proposed
method allows us to classify and localize in the same time the smoke regions with an
interesting computation time and an improved performance metrics compared to CNN based
smoke detection and other aforementioned methods.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, the authors proposed a smoke detection algorithm based on a  Hybrid solution using Deep Belief Network (DBN) classifier. The proposed model seeks to use specific features, including smoke-color, smoke-motion, and energy in order to improve the smoke detection rate. 

For experimental results, the proposed approach has been validated on a larger database containing smoke and no-smoke videos.

The proposed idea is interesting, however, some revisions have to be made and some parts of these experiments are not complete to claim the advantage of the proposed method :

1. The description of the proposed end-to-end network is not clear. Could the authors add more details and explain how their end-to-end model work?
2. The title of the presented paper contains the word « Hybrid solution », but this term not been explained in the manuscript? Could the authors give a clarification about it?
3. In the proposed method, could the authors explain what is the main contribution over standard smoke detection methods based on deep learning, etc ? what are its advantages over them? Could the authors add more clarification?
4. In experimental results, did the authors randomly choose the training samples? What happens when you change the selected frames (another random sampling of training data)?
5. Moreover, what is the motivation to use the SGD optimizer? did the authors test other optimizers like Adam, RMSprop?
6. The pre-training loss curve presented in figure 9, related to the training or test samples? Could the authors add more clarifications?
7. Could the authors add more details and explanation about the Figure 8? It reports a simple architecture of a standard DBN? what is the innovation while the title is « the proposed DBN »
8. Some hyper-parameters are missing, e.g. the size of hidden layers, the size of batch size, etc
9. The English and format of this manuscript should be checked very carefully.

Author Response

Answers to Reviewers 2 comments
Manuscript title: An Efficient Smoke Detection Algorithm Based on Deep Belief Network
Classifier Using Energy and Intensity Features
Authors : Rabeb Kaabi , Moez Bouchouicha, Aymen Mouelhi , Mounir Sayadi, Eric
Moreau
We are very grateful to the Reviewer 2 for his several corrections and recommendations
which greatly improve our manuscript. We agree with all the reviewer’s comments and have
corrected the suggested errors and mistakes and we have done the appropriate changes to
fulfil the comments. Below, we provide specific and detailed responses for each question.
Question 1:
The description of the proposed end-to-end network is not clear. Could the authors add more
details and explain how their end-to-end model works?
Response 1: More details are presented in Lines 24-31
To this end, the selected used features in the preprocessing step in the proposed method
contain the following data: smoke-color, smoke-motion, and energy and allows us to obtain
the smoke regions. Gaussian Mixture Model is employed obviously to capture the frames
containing high motion. After applying RGB rules in smoke pixels and analyzing the energy
attitude of smoke regions, extracted features are then used to fed a DBN for classification.
The contribution of this work lies in the fact that we worked on new combination of smoke
features to extract the regions of interest that we inserted within a simple architecture to
improve the detection rate and localize smoke regions with an interesting computation time
with an improved performance metrics.
Question 2: The title of the presented paper contains the word « Hybrid solution », but this
term not been explained in the manuscript? Could the authors give a clarification about it?
Response 2: the manuscript is untitled "An Efficient Smoke Detection Algorithm Based on
Deep Belief Network Classifier Using Energy and Intensity Features". The word Hybrid is
just a typo in the submission.
Question 3: In the proposed method, could the authors explain what is the main contribution
over standard smoke detection methods based on deep learning, etc ? what are its advantages
over them? Could the authors add more clarification?
Response 3: More details are presented in Lines 237-246.
CNN-based smoke detection methods are significantly slower due to an operation such as
maxpool and the training process takes a lot of time if the computer does not consist of a good
GPU. To this end, the contribution of this work lies in the fact that we worked on the
preprocessing step on new combination of features: smoke-color, smoke-motion, and energy
to extract the regions of interest that we inserted within a simple architecture. Our proposed
method allows us to classify and localize the smoke regions with an interesting computation
time and an improved performance metrics. The proposed method presents a high
performance metrics with an interesting computation time compared to smoke detection
methods based on deep learning (Detection rate, Precision F1 score, accuracy, recall) with the
capability of localization of smoke regions.
Question 4: In experimental results, did the authors randomly choose the training samples?
What happens when you change the selected frames (another random sampling of training
data)?
Response 4: By the way, the choice of tests was based on the fact that the database was
divided into 70 percent for training, 20 percent for validation and 10 percent for testing. For
the training phase, we opted for samples containing smoke and samples containing no smoke
frames and samples containing clouds, fog… .
The most of videos are containing smoke regions. To test the robustness, we tested with
different videos from validation and training.
Question 5: Moreover, what is the motivation to use the SGD optimizer? did the authors test
other optimizers like Adam, RMSprop?
Response 5:
We added a section in the manuscript (Lines 433-456) to detail the choice of the optimizer
with curves and explanations.
We added a comparative part between SGD and Adam optimizer with graphs and explanation
of the used technique of optimization.
Question 6: The pre-training loss curve presented in figure 9, related to the training or test
samples? Could the authors add more clarifications?
Response 6: The pre-training loss curve presented in figure 10 (Figure 9 previously) is related
to training samples
Question 7: Could the authors add more details and explanation about the Figure 8? It reports
a simple architecture of a standard DBN? what is the innovation while the title is « the
proposed DBN »
Response 7: The figure is changed by another explicative figure where we exposed the
architecture of our network in Figure 8.
Question 8: Some hyper-parameters are missing, e.g. the size of hidden layers, the size of
batch size, etc
Response 8: Done in the Table 2.
Question 9: The English and format of this manuscript should be checked very carefully.
Response 9: Done

Author Response File: Author Response.pdf

Reviewer 3 Report

Please see the attached pdf file.

Comments for author File: Comments.pdf

Author Response

Answers to Reviewers 3 comments
Manuscript title: An Efficient Smoke Detection Algorithm Based on Deep Belief Network
Classifier Using Energy and Intensity Features
Authors : Rabeb Kaabi , Moez Bouchouicha, Aymen Mouelhi , Mounir Sayadi, Eric
Moreau
We are very grateful to the Reviewer 3 for his several corrections and recommendations
which greatly improve our manuscript. We agree with all the reviewer’s comments and have
corrected the suggested errors and mistakes and we have done the appropriate changes to
fulfil the comments. Below, we provide specific and detailed responses for each question.
The manuscript addresses an interesting problem encountered in smoke detection and
localization, in which a deep belief network is used as a classifier. The authors applied the
proposed algorithm to 12 videos containing only smoke; clouds, fogs and moving people;
without smoke, and obtained high accuracy and intersection over union metric. A deep belief
network is a mundane technique; however, it is worthwhile to investigate the detection
performance of the algorithm for smoke detection. However, the manuscript contains the
following deficiencies. I suggest that the authors remedy them and submit a revised version. •
Question 1 : The proposed algorithm should be shown as a figure for clarification.
Response 1: We expose in the Figure 5 an explanation in the proposed methodology as a
clarification figure with. Mode details are presented in lines 230-246.
The main idea of our work is to select the frames containing a movement by using GMM. We
eliminate the static frames. We calculate the difference between the components R, G, B of
each frame. For the frames where R, G, B (experimental thresholding) are close we keep
them. We calculate the ratio of the energies of each frame and a reference frame before the
appearance of the smoke. If the contours are blurred, the energy ratio is less than 1.Therfore,
we keep these frame containing smoke and extract the regions of interest. These regions of
interest are inserted then in a vector which will be fed into a DBN network. The proposed
method presents a high performance metrics with an interesting computation time (Detection
rate, Precision F1 score, accuracy, recall) with the capability of localization of smoke regions.
Question 2: It is unclear why equation (5) can detect smoke.
Response 2: More details are given in lines 279-286.
Many studies prove that the decrease in the energy ratio of the current frame E(Bk, It) divided
by the background energy E(Bk, BGt) can be a characteristic of the smoke transparency due
to the fact that smoke gradually softens the edges in an image (blurred contours). Researchers
apply the discrete wavelet transform [16] to the current frame It and the background image Bk
before the appearance of the smoke in order to compute the energy ratio that should be
consequently <1 for smoke regions.
Question 3: Figure 4 Inequalities R - G
Response 3: The frames containing smoke are characterized by closed R, G, B
components. The idea is to set an experimental threshold where R-G, R-B, G-B are close to
zero.
Question 4: Line 240–253: The contents of the paragraph are suitable for the introduction
section.
Response 4: The introduction is modified as recommended by the reviewer.
Question 5: The availability of the datasets employed in this manuscript should be stated.
Response 5: The used dataset is available at [38].

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The answers given by the authors are satisfactory and the manuscript is now ready to publish.

Reviewer 2 Report

The authors have revised this manuscript carefully according to my questions. I have no further questions about this manuscript. It could be accepted.

Reviewer 3 Report

The manuscript addresses an interesting problem encountered in smoke detection and localization, in which a deep belief network is used as a classifier. The reviewer recommends the revised manuscript to be published in the Electronics Journal.