394 Advances in Parallel Computing Algorithms, Tools and Paradigms

D.J. Hemanth et al. (Eds.)

© 2022 The authors and IOS Press.
This article is published online with Open Access by IOS Press and distributed under the terms
of the Creative Commons Attribution Non-Commercial License 4.0 (CC BY-NC 4.0).
doi:10.3233/APC220055

Fire Detection Using Deep Convolution

Neural Network in Video Streams
P.M.Ashok kumara,1, Lakshmi HarikaPalivelab and V.V. Rama Krishnac
a
Department of CSE, KoneruLakashmaiah Education Foundation, Vaddeswaram, India
b
Department of IT, MIT, Anna University, India
c
Department of ECE, Lakireddy Bali Reddy College of Engineering,Mylavaram, India

Abstract.The concept of Convolution Neural Networks (CNN) has been becoming

highly significant in computer-vision-based applications. Their applications in
disaster management, like fire detection, will definitely improve the social and
ecological environment. Most of the existing fire detection systems fail to detect
fire in certain environments like smoke, fog and so on. In this paper, we propose a
Squeeze-Net framework-based CNN for detection of fire, localization and
understanding the scene of fire. The method uses smaller convolution layers with
no dense layers, thus minimizing the computational power. Experimental results
suggest improved performance in terms of accuracy and loss parameters for both
known and unpredictable image settings. Despite the low computational power,
the method provides more accuracy than state-of-art techniques.

Keywords. Convolution Neural Network, Feature Map, Fire, Squeeze Net.

1. Introduction

A variety of sensors has been developed for various applications, like diagnosis of
diseases, navigation monitoring, abnormality detection, suspicious behavior detection,
fire detection [1] and so on.Using smart video surveillance systems, a variety of
abnormalities [2] like fire, smoke, fog can be easily detected. In recent days, the
Amazon Forest fire has damaged the great diversity of our lives. In order to deal with
these disasters, these abnormalities should be dealt with in earlier stages and detected
to avoid dangers. Abnormality detection [3,4] has now become a great challenge.
In earlier days, Fire [6] could be detected in two ways: using traditional alarm
systems and sensor assisted systems. The traditional security [7] devices rely on
thermal and visual detectors. The traditional alarm system needs a human interface to
check if the fire is present and needs close monitoring. To address the limitations of
traditional systems, the sensor-based system has been developed in order for faster
response. The sensor-based systems, however, don’t need human interface and they
provide additional details about the location and severity of fire. But it failed in the
case of poor lighting and low-quality frames.

1
P.M.Ashokkumar, Department of Computer Science and Engineering, Koneru Lakashmaiah Education
Foundation, Vaddeswaram, India. Email: [email protected]
 P.M. Ashok Kumar et al. / Fire Detection Using Deep Convolution Neural Network 395

2. Related Works

Sathyavathiet al.,[5] suggested an image processing-based system for abnormality

detection from video streams. An RGB model-based hue and disturbance assessment
was utilized to extract fire pixels and smoke pixels. Marbach et al [14] used chromance
space to determine non-fire/blaze regions.
Dimitropoulos K et al.,[9] proposed a dynamic spatio-temporal flame model, in
which prior assumptions of the fire occurrence in neighboring blocks are used to
estimate the spatial-temporal consistency of the fire zone in order to increase resilience.
Celik et al.,[5] suggested a generic chromance model. Foggia et al. [11] suggested a
method based on color and form, in which the system incorporates real-time fire
detection as well as multi-expert systems. The downside is that the technique is
sensitive to variations in brightness, resulting in a huge number of false positives owing
to the existence of shadows.
Filonenko et al.,[10] developed a CUDA-based Quick Smoke Detection Scheme
for Surveillance Cameras. The probability that pixels belong to smoke is assessed by
the color of the smoke. But the main limitation is speed may be a bottleneck between
the CPU and GPU.
Di Lascio R et al.,[8] developed a strategy for increasing fire detection
effectiveness through the use of advanced analytics. The experimental survey on
literature shows that computational methods give better loss rates and methods that are
simple, whereas they compromise accuracy rates. Deep CNN [12,15] has been
proposed for fire detection to address the aforementioned difficulties. The main
contributions are outlined below.
 Time-consuming efforts have been eliminated, and the suggested technique
delivers more accuracy with lower false alarm rates than state-of-the-art
systems.
 Our model has been fine-tuned and trained using deep CNN, which has dense
layers, and initially used to classify the frames with fire. This will give
accuracy related to the previous models.
 In order to improve accuracy, the Squeeze Net model has been used. Squeeze
Net is not only known for its accuracy but also for its less computational
power and less memory.
The remaining sections are as follows. Section 3 is concerned with the intended
work and architecture. The datasets and experimental findings are discussed in Section
4. In Section 5, the paper is concluded and future works are discussed.

3. Proposed Work

Earlier approaches of fire detection systems such as the YUV model based on spatial
and temporal wavelet transform for categorization into fire and non-fire. These color
models were only used for static fires. In subsequent phases, multi-wavelet analysis
was used to improve the threshold in the RGB color model, lowering the false alarm
rates. The SVM classifier predicts fire video streams at the expense of increased
computational complexity. In order to overcome the mentioned drawbacks, the CNN
model has been induced in scene for efficient fire detection. The architecture diagram
for the proposed work is shown in Figure 1.
396 P.M. Ashok Kumar et al. / Fire Detection Using Deep Convolution Neural Network

Figure 1. Fire localization using proposed Deep CNN Figure 2 Classification of fire frames
and labeling

3.1. Convolution neural network

The proposed CNN model uses different filters in the convolution layer and pooling
layer to produce different compact feature maps. We use the Squeeze net model for
evaluation of the confidence percentage of the fire. The Squeeze Net is known for its
accuracy and low computational power. It gives more accurate results when compared
to other training models. It down samples the input channels so that the convolution
layer has larger activation maps. Fire modules are the building blocks of the Squeeze
Net model. It has two layers that pinch and expand.A Squeeze net is comprised of a
stack of fire modules and a few pooling layers. The feature maps in the squeeze and
expanded layers are the same size. The squeeze layer decreases depth while the
expanded layer raises it.
The squeeze net model architecture has been trained with around 1000 classes. But
in the proposed method, only two classes are needed. Thus, only two classes, fire and
normal, are considered. The architecture consists of a convolution layer followed by
max-pooling. The input of fire modules is of size 224*224*3. Every fire system has
two convolutions, one for squeezing and one for development. There is a max pooling
layer after the first fire alarm modules. After the last fire module, the model is modified
accordingly to obtain two classes. A transfer learning strategy has been used for
training the Squeeze Net model. Thus, accuracy was increased to 5 percent more than
previously. The model was trained for 10 epochs and an accuracy of around 98 percent
has been achieved. Transfer learning is employed to solve computational problems
efficiently.

3.1.1. Deep Convolution neural network

Deep CNN models automatically learn powerful features from datasets. Another
motivator is elimination of preprocessing and pattern discovery, both of which are
required by fire hazard detection. The frames are trained based on two classes: fire and
 P.M. Ashok Kumar et al. / Fire Detection Using Deep Convolution Neural Network 397

normal. If the frames are non-fire, then, no further process is required. If the frame is
classified as fire, then further feature extraction is done, followed by binarization.The
classifications are allocated based on the network's probability ratings. The higher the
likelihood score, the higher the input image's final class value. Figure 2 depicts a
succession of images with classifications.

Table 1. Feature Map CalculationAlgorithm

Input: Given Training_Samples (Ts), Ground_Truth(gt),
CNNmodel (CNNM)
1.Start
2.Forward propagate training samples through CNNM
3. Selection of feature maps from the layers
4. Resizing the ground truth and Feature maps to 256×256
pixels
5. Compute activation mean for Feature maps
6. Feature map binarization as follows:
F(x,y) bin(i) = 1, F(x, y)i> FMA(i) (1)
= 0,Otherwise
7. Compute the hamming distance between the actual truth and
each feature space using the formula below:
HDi = (Fbin(i) – gt) (2)
8.Calculate the sum of all resultant distances, and shortlist the
minimum distances using threshold T
9. Choose relevant feature maps based on the hamming
distances you've narrowed down.
10.End

3.2. Feature map calculation

The intermediate layers are useful in understanding how successive CNN layers
transform their input when moving from one layer to another layer. The feature maps
are called activations. The feature maps help in understanding the neural network, how
the network is trained and how the input is decomposed by each layer.
The features are assessed using the feature map selection technique, and the layers
that are determined to be very sensitive to wildfire zones. After the forward
propagation of datasets through the CNN network, feature maps are selected from any
layer on CNN and mean activation maps are calculated. Binarize means activation
maps are obtained and hamming distance is calculated. A minimum of hamming
distance is chosen by setting a threshold and thus feature maps sensitive to fire will be
obtained. The initial layers retain the original shape and features of the frame or image.
The first layer retains almost all information about the input image. While traversing
through the dense layers, the features become increasingly abstract. The higher layers
retain only less information about the frame and is less visually predictable. The last
layers will not activate at all, there is nothing to learn for a CNN model from those
layers.

4. Experimental Results

The experimental findings on the dataset were described, followed by a comparison

with earlier models.
398 P.M. Ashok Kumar et al. / Fire Detection Using Deep Convolution Neural Network

4.1. Experiment on datasets

General benchmark datasets have been used for experimental analysis. Datasets were
collected from MNIST sites. Another set of FIRE-DATASETS were collected and used
for our approach. The datasets have been split into two forms: training and testing.
Fifty percent of the datasets were utilized for training and the remaining for testing.
The model was trained using both deep CNN and the squeeze net model.

4.2. Accuracy Percentage

The datasets were trained for the deep CNN model and tested the Squeeze Net model.
When tested with the Squeeze Net model, more accuracy was achieved than the
previous model, Alex net [13]. Table 2 shows a comparison of accuracy percentages.
When contrasted to prior approaches, our suggested technique had an accuracy of
around 94 percent, which is 5% higher than previous methods. Table 3. results
performance values.
The false alarm rate is reduced from 9.07% to 8.06%. The suggested work's training
accuracy graph is displayed in Figure 3 and has an accuracy of roughly 94 percent.

4.3. Confidence score

When tested with the Squeeze Net model, the confidence score achieved is shown in
Figure 4. The confidence score is about the surety of fire in the frame along with
classification. The confidence score absorbed for frames with intense fire will be
around 99%.

Figure 3Training accuracy plot. Figure 4. Training Loss plot

Table 2. AccuracyComparison TABLE 3.Performance Parameters of Proposed Approach

Method Accuracy Percentage Epochs Sensitivity Accuracy False Alarm
1 0.9 0.902 0.1452
Squeeze net 94.04
2 0.9153 0.92 0.1557
Alex net [13] 90.06 3 0.9343 0.93 0.1302
4 0.90 0.91 0.199
Foggia et al [11] 93.55
5 0.9354 0.94 0.1205
Celik et al [3] 83.87 TOTAL 0.9112 0.9304 0.1512
 P.M. Ashok Kumar et al. / Fire Detection Using Deep Convolution Neural Network 399

5. Conclusion

Intelligent CCTV surveillance systems have emerged as a result of intelligent cams

inherent process technology. Inspired by CNN's huge opportunities, this paper offers a
light-weight CNN based on the Squeeze-Net model to detect fire hazards on CCTV
networks. This method is capable of both blaze localization and recognizing the target
under observation. Furthermore, the suggested system uses fine-tuning and the
Squeeze-Net architecture to balance the precision of fire detection with the volume of
fire. Experimental results also suggest that the model performs better than other state
of-art techniques.

References

[1] Adel H et al.,“Investigation of a Novel Image Segmentation Method dedicated to Forest Fire
Applications,” Meas. Sci. Technol., 2013, vol. 24, no. 7.
[2] Ahmad J, Muhammad K, Bakshi S, and Baik S. W, "Object-Oriented Convolutional Features for Fine-
Grained Image Retrieval in Large Surveillance Datasets," Future Generation Computer Systems, 2018,
vol. 81, pp. 314-330.
[3] Celik T and Demirel H, “Fire Detection in Video Sequences Using AGeneric Color Model”, Fire safety
journal, 2009, vol. 44, pp. 147-158.
[4] Chan T.-H et al., “PCANet: A Simple Deep Learning Baseline for Image Classification”, IEEE Trans.
Image Process., 2015, vol. 24, no. 12, pp. 5017–5032.
[5] Sathiyavathi, V., M. Jessey, K. Selvakumar, and L. SaiRamesh. "Smart Surveillance System for
Abnormal Activity Detection Using CNN."Advances in Parallel Computing Technologies and
Applications 40 (2021): 341.
[6] Chi R, Z.-M. Lu, and Q.-G. Ji, “Real-Time Multi-Feature-Based Fire Flame Detection in Video,” IET
Image Process.,2016, vol. 11, no. 1, pp. 31–37.
[7] Chino D. Y. T, Avalhais L. P. S, Rodrigues J. F et al., “BoWFire: Detection of fire in still images by
Integrating Pixel Color and Texture Analysis,” in Proc. 28th SIBGRAPI Conf. Graph. Patterns Images,
2015, pp. 95–102.
[8] Di Lascio R, Greco A, Saggese A, and Vento M, "Improving fire detection reliability by a combination
of videoanalytics," in International Conference Image Analysis and Recognition, 2014, pp. 477-484.
[9] Dimitropoulos K, Barmpoutis P, and Grammalidis N, "Spatio-Temporal Flame Modeling and Dynamic
Texture Analysis for Automatic Video-based Fire Detection," IEEE transactions on circuits and
systems for video technology, 2015, vol. 25, pp. 339-351.
[10] Filonenko A, Hernández D. C, and Jo K.-H, “Fast Smoke Detection for Video Surveillance Using
CUDA”, IEEE Transactions on Industrial Informatics, 2018, vol. 14, pp. 725-733.
[11] Foggia P, Saggese A, and Vento M, “Real-time fire detection for video-surveillance applications using
a combination of experts based on Color, Shape, And Motion”, IEEE TRANSACTIONS on circuits
and system for video technology, 2015, vol. 25, pp. 1545-1556.
[12] Han D and Lee B, “Development of early Tunnel Fire Detection Algorithm using the image
processing”, in International symposium on visual computing, 2006, pp. 39-48.
[13] Iandola F. N et al., “Squeeze Net: Alex-Net-level accuracy with 50x fewer parameters and <1MB
model size,”, 2016.
[14] Marbach G, Loepfe M, and Brupbacher T, “An image processing Technique for Fire Detection in
Video Images”, Fire safety journal, 2006, vol. 41, pp. 285-289.
[15] Muhammad K, Ahmad J, Lv Z, Bellavista P, Yang P, and Baik S. W, "EfficientDeep CNN-Based Fire
Detection and Localization in Video Surveillance Applications," IEEE Transactions on Systems, Man,
and Cybernetics: Systems, 2018, pp. 1-16.