EAST WEST INSTITUTE OF TECHNOLOGY

AN IMPROVED APPROACH FOR FIRE DETECTION USING

DEEP LEARNING MODELS

PRESENTED BY: GUIDED BY:

B G Sumith Kumar MR Hemanth Kumar
Asst. prof of Dept of ISE, EWIT
 CONTENTS

• Introduction
• Existing work
• Proposed work
• Implementation
• Results
• Conclusion
• References
 INTRODUCTION

Fire is the big threat in human life like floods, and earthquakes. Situation like fire causes threat
to public safety and health. It is the more commonly occurring abnormal event than other
abnormal events such as earthquakes and floods. Fire detection system, which are made of
hardware, are quite expensive. Hence, researchers proposed image processing and computer
vision-based techniques for fire detection. Fire detection using computer vision techniques and
image processing has been a topic of interest among the researchers. Indeed, good accuracy of
computer vision techniques can outperform traditional models of fire detection. However, with
the current advancement of the technologies, such models of computer vision techniques are
being replaced by deep learning models such as Convolutional Neural Networks (CNN).
 EXISTING WORK

• Fire could have the most dynamic features such as area of the flame, randomness of the
flame.
• Due to this reason, feature extraction using computer vision technique made this process
very difficult.
• With the enhancement of the technologies, deep learning concept has been introduced.
• CNN based deep learning model is the most commonly used model in detection of fire.
• Performance of CNN has been found to be quite better as compared to computer vision
techniques.
• Due to lack of large datasets, sometimes CNN does not give satisfactory results.
 PROPOSED WORK

• The proposed approach in this paper uses Deep CNN instead of traditional CNN.
• Deep CNN is based on the transfer learning that uses pre-trained model to train another
model, which is being trained to detect different object.
• Two Deep CNNs, VGG 16 and MobileNet have been used, which outperform traditional
CNN model. The models are trained on ‘imagenet’ dataset.
• Proposed approach architecture consists of four phases and working of these phases is
described in the following sub-sections.
Data Preparation

• The data consist for the proposed work consists of images in jpg format. These images
are handpicked through various sources over the internet.
• After processing, dataset is labelled into training and testing data. Both training and
testing datasets are prepared by sampling frames into ‘default’ and ‘fire’ categories.
Frames are kept in directories named after the class.
Model creation

• The proposed work has been implemented by using two Deep convolution neural
networks, i.e., VGG16 and MobileNet.
• VGG16 uses sixteen dense layers of neural networks, while MobileNet uses depth-wise
separable convolution to create model.
• These models are created by using convolution layer of different channels with
activation function and performing max pooling on these layers.
• The final model is created by adding numb er of ‘Dense’ layers with number of
activation units equivalent to number of classes.
• Created models act as an input for the training the model along with the datasets.
Model Training

• To train model data augmentation is done by performing pre-processing function in

pre-processing input. The data for training and testing are read from directories using
data generator function.
• After pre-processing training and testing data, these data send to model fitting for
model training.
Final Result

• Final result shows different graphs of accuracy and loss of training and testing data with
epochs (go over training data epoch times and each time update results).
• Convolution Neural Network (CNN) model table is drawn as well as Deep Convolution
Neural Networks tables.
• Final result is tested whether image belongs to fire or non-fire category.
 IMPLEMENTATION

VGG16

• The architecture of the VGG was introduced by Karen Simonyan and Andrew Zisserman.
• It is based on transfer learning which uses deep neural networks to implement.
• It is possible to use the learned weight in one model to be transferred in another model
with the help of deep neural networks model.
• For example, if a model that is used to detect dogs and cats in images can be used in
detection of fire and smoke.
• VGGNet are classified into VGG16 and VGG19.
Architecture of VGG16
MobileNet

• The MobileNet architecture is based on the depth-wise separable convolutions.

• Depth-wise convolutions mean that it performs a single convolution on each color
channel rather than flattening it.
• Depth-wise convolution is unlike spatial separable convolution.
• It cannot be factored in two or smaller kernel size.
• Depth-wise separable convolution deals with both spatial dimensions and depth
dimensions as well.
• The proposed work uses MobileNet as it is lightweight in nature.
MobileNet Architecture
 RESULTS
• The performance of fire detection is tested against traditional CNN, and two deep
learning models VGG16 and MobileNet using transfer learning to increase accuracy.
However, this may also increase training time too.

Epoch Steps Loss Accuracy Val_loss Val_acc

1 88 0.5920 0.6922 0.7304 0.6960

16 88 0.2505 0.9141 0.1714 0.9284

49 88 0.2240 0.9053 0.4713 0.9341

50 88 0.2611 0.9112 0.3337 0.8968

sample results of CNN model

• Epoch=50, specifies to go over training data 50 times and each time updates are made
and Steps=88, means that how many images are to be processed before making a
gradient update. It is calculated by dividing total images to the batch size.

Value of loss function with epochs using CNN

 Epoch Steps Loss Accuracy Val_loss Val_acc

1 88 1.3413 0.8785 0.5258 0.9463

25 88 0.1361 0.9915 0.3033 0.9573

50 88 0.1360 0.9915 0.3143 0.9573

sample results of VGG16 model

. Value of loss function with epochs using VGG16

 Epoch Steps Loss Accuracy Val_loss Val_acc

1 88 0.2496 0.9221 0.6546 0.8793

4 88 0.0781 0.9745 0.1387 0.9584

10 88 0.0424 0.9848 0.1055 0.9570

sample results of MobileNet model

Value of loss function with epochs using MobileNet

 CONCLUSION

• The proposed work uses both Convolution Neural Networks and Deep CNN using transfer
learning for the detection of fire in images.
• Convolution Neural Network (CNN) was tested on our dataset and accuracy of training and
validation was observed and we have plotted graphs of loss and accuracy in accordance with
epochs.
• After CNN model we have used Deep CNN models for testing on our dataset and accuracy of
Deep CNN models was observed and we have plotted different graphs of loss and accuracy in
accordance with the epochs.
• We have compared CNN models with VGG model and MobileNet model. Deep CNN uses
deep transfer learning approach to train the model that’s why these give better performance.
• As we observed in the graph of training versus validation accuracy that Deep neural networks
have problem of overfitting the training data. There is still some work that needs to be done.
 REFERENCES
• R.Bright and R.Custer, “Fire detection: The state of the art,” NBS Technical Note, US Department
of Commerce, 1974.
• V.Vipin, “ Image processing-based forest fire detection,”International Journal of Emerging
Technology and Advanced Engineering, 2.2 (2012), pp. 87-95.
• P.V.K. Borges, E. Izquierdo,”A probabilistic approach for visionbased fire detection in videos,”
IEEE Trans. Circuits Syst. Video Technol, 20.5 (2010), pp. 721–731.
• S. Frizzi, R. Kaabi, M. Bouchouicha, J. M. Ginoux, E. Moreau, and F. Fnaiech, “Convolutional
neural network for video fire and smoke detection,”In IECON 2016 42nd Annual Conference of the
IEEE Industrial Electronics Society, Oct 2016, pp. 887-882.
• S. Hochreiter,J. Schmidhuber, ”LSTM can solve hard long time lag problems,” In Advances in
Neural Information Processing Systems; NIPS: San Diego, CA, USA, 1997; pp. 473–479.
• S. Kethavath, and M. Dua. "Early Discovery of Disaster Events from Sensor Data Using Fog
Computing," International Conference on Intelligent Computing, Information and Control Systems.
Springer, Cham, 2019.
THANK YOU