Flame Detection Using CNN in Safety Surveillance

Abstract – In tradition flame detection system, we generally use cameras, sensors and circuits in a stable environment. Usage of this tradition methods are the means to loss, and more over consumes more computational cost to detect fire. To overcome this loss, we use Convolutional neural network (CNN). In this we are using Squeeze Net architecture for flame detection, localization methods to image classi?cation fire and non-fire frame. Morphology process will be very useful for this purpose which process images based on shapes.

After this we will apply filtrations to reduce noises in frame.

The result of this process yields to identify the flame very ef?ciency and more over the accuracy will be more than 95%.

Keywords- Deep learning, Convolutional neural networks (CNNs), ?re detection, ?re disaster analysis, ?re localization, image classi?cation, surveillance networks.

I. INTRODUCTION

Traditional system usually involves more of embedded system and circuities and more of human involvement thus requires more computational cost .the usage of CCD camera in surveillance thus not suitable for all kind of environment, during fire detecting in incident or disaster basically in traditional method to detect fire range wave length and spreading speed involves human to go nearby and to see the fire range during this there maybe human loss and may also cause harm to the human who are near by the fire range.

However, the collected information turns to be wrong as the fire keeps spreading. To overcome all this, we go for visualization image processing. For this process CNN will be very helpful.

Get quality help now

Writer Lyla

Verified

Proficient in: Government

5 (876)

“ Have been using her for a while and please believe when I tell you, she never fail. Thanks Writer Lyla you are indeed awesome ”

+84 relevant experts are online

Hire writer

Major advantages are less human interference, faster response, affordable cost, and larger surveillance coverages. It also provides other information like ?re including its location, size, degree, etc.

By using this CNN modules, we can able to detect the fire in early stage and can reduce the damage to minimal.

Training and ?ne-tune of Alex Net architecture for ?re detection can be done using a transfer learning strategy. Thus, its size is implementation dif?cult in resource constrained equipment.

To reduce the size of implementation we will use Squeeze Net model for ?re detection at the early stages (238 to 3 MB). So, it is feasible to the cost and proposed requires 0.72 GFLOPS/image compared to Alex Net, whose computational complexity is 2 GFLOPS/image.

On the bases of feature selection map, it can be able to detect fire region early. Using segmentation information and analyses of segmentation we can detect the fire more accurate.

Fig 1.Overview of the proposed system for ?re detection using CNN.

Survey papers- Zhenglin Li, Lyudmila S. Mihaylo, in Autonomous Flame Detection in Videos With a Dirichlet Process Gaussian Mixture Color Model here frame accuracy higher than 95%. And there will be no ?ltering out the disturbance in some areas.

Chi Yuan, Zhixiang Liu and Youmin Zhang in Fire Detection Using Infrared Images for UAV-based Forest Fire Surveillance uses histogram based segmentation and optical ?ow analysis, ?res from background as well as non-?re hot moving objects can be well differentiated and can improved performance. CCD cameras reliable enough in some outdoor applications.

Wonjae Lee, Seonghyun Kim, Yong-Tae Lee in Deep Neural Networks for Wildfire Detection with Unmanned Aerial Vehicle here system achieved high accuracy for wide range of aerial photographs. And it suffers from low accuracy and a long training time because of over fitting and a large amount of parameters.

The process is carried out in the following steps :

Developing the Squeeze Net model.

Training model for classification on fire and non-fire datasets

Video will be converted into frame

Reducing size of Frame and fitting to the module

Identifying the candidate region

Applying filtration to reduce the noise

Finding candidate region and identifying the fire frame , light conditions, objects like fire

II. PROPOSED FRAMEWORK

Fire detection using image processing is a time taking task, due to the time-consuming for methods of image Processing. Detecting fire at early stage is challenging task because of similar fire like objects. It also provide high false alarms and low accuracy. Deep learning models provide good results on fire detection on early stage.

Convolutional Neural Network Architecture

Convolutional neural network is one of the most used model for image classifications, recognition.

CNNs are consisted of input layer, output layer, hidden layer. hidden layers of a CNN consist of convolutional layers, RELU layer, pooling layers, fully connected layers. 8 CNN layers have been used.

Each input image will pass it through a series of convolution layers with filters ,Pooling, fully connected layers and apply Softmax function to classify an fire and non fire with probabilistic values between 0 and 1.

1. Convolution Layer- First layer in CNN is Convolution layer it extract features from an input image dataset. Convolution perform operation of pixels to extract image features using small squares of input data, process the data and send it to next layer.

2. Pooling layers- Pooling layers section would reduce the dimensions when the images are too large. Spatial pooling also called sub sampling which reduces the dimensions of each map but retains the important information. Max pooling take the largest value element from the feature map. Taking the largest element could also take the max pooling. ReLU stands for Rectified Linear Unit for a non-linear operation. The output is

?(x) = max(0,x).

3. Fully Connected Layer-The layer we call as FC layer, we flattened our matrix into vector and feed it into a fully connected layer like neural network.

III. EXPERIMENTAL RESULTS

Data Material- We have used fire and non fire dataset which contains 651 images of both fire and non fire.

Fig 2.set of sample images of fire and non fire

We conducted the experiments using a system with the below speci?cations: Memory 2GB GPU NVIDIA GeForce GTX650 and windows OS installed on an Intel Core i5 CPU with 6 GB RAM. A total of 651 images.

For the training phases of the experiments, we followed the experimental strategy, where 80% and 20% of the data are used for training and validation respectively. we trained our proposed SqueezeNet model with some more real-time dataset for getting high accuracy.

For the testing phases we getting real-time video through camera and also testing video for checking fire or non fire on each frame in video.

We evaluated the images of the test set with the CNN model.The number of BatchSize is 10, the validation split is 2/10 and epochs is 20. The accuracy rate reached above 90%

We started training by 2 classes to the model with 5 epochs and 32 batch size.Accuracy rate was reached upto 80%

Input Images Segmented fire Output images Category

Fire

Fire

Fire

No Fire

No Fire

No Fire

Fig 3.Sample outputs from our system: the ?rst column shows input images, the second column shows segmented fire images, the third column highlights the results for each image

By training by 2 classes to the model with 20 epochs and 10 batch size.Accuracy rate was reached more than 95%

Here we can observe, model has attained the accuracy above 90% when model trained with more epochs and less batch size. This method provides high accuracy and takes more time to train network

IV CONCLUSION

In this paper the ‘Flame Detection Using CNN in Safety -Surveillance ‘we are going to detect flame frame using CNN networks.

The main advantage of using CNN helps as to identify the fire frame faster than the tradition, as tradition involves more human interface. disasters can be stopped at the early stages and can control human and environment losses.

Furthermore, it has few disadvantages as if we install it in a new environment it detects non-fire also as fire, to overcomes this we need to train the environment first and test it later.

It requires more training data to achieve higher accuracy.

In future using superior algorithms we can still ignore the human interface and make the process more automated.

REFERENCES:

I. Mehmood, M. Sajjad, and S. W. Baik, “Mobile-cloud assisted video summarization framework for ef?cient management of remote sensing data generated by wireless capsule sensors,” Sensors, vol. 14, no. 9, pp. 17112–17145, 2014.

II. T. Toulouse, L. Rossi, M. Akhlou?, T. Celik, and X. Maldague, “Benchmarking of wildland ?re colour segmentation algorithms,” IET Image Process., vol. 9, no. 12, pp. 1064–1072, Dec. 2015.

III. J. Ahmad, K. Muhammad, S. Bakshi, and S. W. Baik, “Object-oriented convolutional features for ?ne-grained image retrieval in large surveillance datasets,” Future Gener. Comput. Syst., vol. 81, pp. 314–330, Apr. 2018.

IV. Yang C, Yuan J, Ji L. Sparse reconstruction cost for abnormal event detection[C]// IEEE Conference on Computer Vision & Pattern Recognition. 2011:3449-3456.

V. Liu C B, Ahuja N. Vision Based Fire Detection[C]// International Conference on Pattern Recognition. IEEE, 2004:134-137.

VI. Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015.

VII. “Tensor?ow,” accessed: 2017-03-15.

VIII. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classi?cation with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.

IX. J.Yang, X.Yuan, X.Liao, P.Liul, G.Sapiro, L.Carin“ Video Compressive Sensing Using Gaussian Mixture Models,” IEEE Transactions on Image Processing, Volume: 23, Issue: 11, Nov. 2014

X. R.Kaabi, S.Frizzi, M.Bouchouicha, F.Fnaich, E.Moreau, “Video Smoke Detection Review, ”International conference IEEE on Smart Monitored and controlled Cities SM2C, 2017.

XI. Y.WANG, “Smoke Recognition Based on Machine Vision,”2016 International Symposium on Computer, Consumer and Control (IS3C) (2016),July 2016.

XII. C. Durmaz, M. Challenger, O. Dagdeviren and G. Kardas, “Modeling Contiki-based IoT Systems,” 17th Symposium on Languages, Applications and Technologies, 26-27 June 2017.

XIII. S. Arslan, M. Challenger, O. Dagdeviren, “Wireless Sensor Network based Fire Detection System for Libraries” 2nd International Conference on Computer Science and Engineering, Shanghai, 17-2017

XIV. P. Foggia, A. Saggese, and M. Vento, “Real-time ?re detection for videosurveillance applications using a combination of experts based on color, shape, and motion,” IEEE Trans. Circuits Syst. Video Technol., vol. 25, no. 9, pp. 1545–1556, Sep. 2015.

XV. M. Mueller, P. Karasev, I. Kolesov, and A. Tannenbaum, “Optical ?ow estimation for ?ame detection in videos,” IEEE Trans. Image Process., vol. 22, no. 7, pp. 2786–2797, Jul. 2013.

XVI. T.Toulouse,L.Rossi,T.Celik,andM.Akhlou?,“Automatic?repixeldetection using image processing: A comparative analysis of rule-based and machine learning-based methods,” Signal, Image Video Process., vol. 10, no. 4, pp. 647–654, 2016.