Flame Detection Using Appearance-Based Pre-Processing and Convolutional Neural Network
Round 1
Reviewer 1 Report
In this study, HSV color conversion and Harris Corner Detection are used in the image pre-processing step to reduce the incidence of false detection. In
addition, among the detected corners, the vicinity of the corner point facing the upper direction was
extracted as a region of interest (ROI) and the fire was determined using a convolutional neural
network (CNN). Some comments for improving the quality of the paper are listed below.
- Figure 6 is not well explained. Why is the design as such?
- The second and third paths result in the same size, but the first path does not.
- Can Table 1 be explained in more detailed? Especially the relationship between the input size and kernel size.
- Please show recall values in Table 4.
- Why is the HSV color model chosen?
Author Response
Thank you for your review,
1. Figure 6 is not well explained. Why is the design as such?
Answer : We have supplemented your comment. (You should refer to line 200-205 on page 7.)
2. The second and third paths result in the same size, but the first path does not.
Answer : In the Inception module, the features are effectively extracted through convolutional kernels of different sizes. The second and third paths are the presence or absence of a 1x1 convolutional layer, and there is a difference in the size of the parameter in the color channel dimension. Two times 3x3 kernel in a row of the first paths is instead of the conventional 5x5 kernel. Therefore more features can be extracted with fewer parameters.
3. Can Table 1 be explained in more detailed? Especially the relationship between the input size and kernel size.
Answer : We have supplemented your comment. (You should refer to line 194-200 on page 7.)
4. Please show recall values in Table 4.
Answer : We have supplemented your comment. (You should refer to Table 4.)
5. Why is the HSV color model chosen?
Answer : When using the HSV color model, the detection range according to the lighting change can be determined by adjusting Saturation and Value.
Therefore the HSV color model is more advantageous than RGB color model to detect objects that exist in reality.
That's all, thank you,
Reviewer 2 Report
The authors discussed in the manuscript how to use image pre-processing techniques to reduce the false detection of fires. While the pre-processing techniques (HSV and Harris corner detection) are well-known in the field, the proposed algorithm to apply these techniques in fire detection shows significant improvement in performance compared to other methods (R-CNN and SSD). However, the authors need to address the follow questions before the manuscript is recommended for publication.
- (Major) One of the most important aspect in fire detection is whether fires can be detected in real-time. The authors need to demonstrate the proposed algorithm can be used in real-time fire detection (i.e., how fast the detection speed can reach, 1fps or 10fps?)
- Is the training/testing dataset publicly available? How representative are the images in the training/testing dataset. The authors need to clarify how they construct the dataset or where the dataset originally come from.
- The authors need to discuss whether they can further improve the performance if data pre-processing is coupled with other object detection methods discussed in the manuscript.
Author Response
Thank you for your review,
- This study is sufficiently available as real-time fire detection. The information related to detection speed is shown on page 12. Based on the experimental hardware, the inference speed per frame is 0.38 seconds. (You should refer to line 282-288 on page 12.)
- The used training/testing image dataset is obtained from published materials for use in research, and the dataset is shown in Table 3. In addition, all the photos used for performance evaluation are taken directly by ourselves.
- Reviewer comment : The authors need to discuss whether they can further improve the performance if data pre-processing is coupled with other object detection methods discussed in the manuscript.
Answer : We have supplemented your comment. (You should refer to line 288-291 on page 12.)
Round 2
Reviewer 1 Report
My previous comments have been satisfactorily addressed in this revision. Thanks.
Author Response
Thank you very much for your comment,
