Skip to content
real-time fire detection in video imagery using a convolutonal neural network (deep learning) - from our ICIP 2018 paper (Dunnings / Breckon)
Python Shell
Branch: master
Clone or download

Experimentally Defined Convolutional Neural Network Architecture Variants for Non-temporal Real-time Fire Detection

Tested using Python 3.4.6, TensorFlow 1.13.0, tflearn 0.3 and OpenCV 3.3.1 / 4.0.x

(requires opencv extra modules - ximgproc module for superpixel segmentation)


FireNet                                                                             FireNet architecture (above) InceptionV1-onFire                                                                           InceptionV1-OnFire architecture (above)


"In this work we investigate the automatic detection of fire pixel regions in video (or still) imagery within real-time bounds without reliance on temporal scene information. As an extension to prior work in the field, we consider the performance of experimentally defined, reduced complexity deep convolutional neural network (CNN) architectures for this task. Contrary to contemporary trends in the field, our work illustrates maximal accuracy of 0.93 for whole image binary fire detection (1), with 0.89 accuracy within our superpixel localization framework can be achieved (2), via a network architecture of significantly reduced complexity. These reduced architectures additionally offer a 3-4 fold increase in computational performance offering up to 17 fps processing on contemporary hardware independent of temporal information (1). We show the relative performance achieved against prior work using benchmark datasets to illustrate maximally robust real-time fire region detection."

(1) using InceptionV1-OnFire CNN model (2) using SP-InceptionV1-OnFire CNN model

[Dunnings and Breckon, In Proc. International Conference on Image Processing IEEE, 2018]

Reference implementation:

Our binary detection (FireNet / InceptionV1-OnFire) architectures determine whether an image frame contains fire globally, whereas the superpixel based approach breaks down the frame into segments and performs classification on each superpixel segment to provide in-frame localization.

This respository contains the and files corresponding to the two binary (full-frame) detection models from the paper. In addition the file corresponds to the superpixel based in-frame fire localization from the paper.

To use these scripts the pre-trained network models must be downloaded using the shell script which will create an additional models directory containing the network weight data (on Linux/MacOS). Alternatively, you can manually download the pre-trained network models from and unzip them to a directory called models in the same place as the python files.

The superpixel based approach was trained to perform superpixel based fire detection and localization within a given frame as follows:

  • image frame is split into segments using SLIC superpixel segmentation technique.
  • the SP-InceptionV1-OnFire convolutional architecture, trained to detect fire in a given superpixel segment, is used on each superpixel.
  • at run-time, this SP-InceptionV1-OnFire, network is run on every superpixel from the SLIC segmentation output.

Which model should I use ? : for the best detection performance (i.e. true positive rate) and throughtput (speed, frames per second) use the FireNet model (example:; for a slighly lower false alarm rate (i.e. false positive rate, but only by 1%) but much lower throughtput (speed, frames per second) use the InceptionV1-OnFire model (example:; for localization of the fire within the image use the superpixel InceptionV1-OnFire model (example: For full details see paper - [Dunnings and Breckon, 2018]

Training datasets:

Original frame (left), Frame after superpixel segmentation (middle), Frame after superpixel fire prediction (right)

Instructions to test pre-trained models:

To download and test the supplied code and pre-trained models (with TFLean/OpenCV installed) do:

$ git clone
$ cd fire-detection-cnn
$ sh ./
$ python models/test.mp4
$ python models/test.mp4
$ python models/test.mp4

Instructions to use pre-trained models with other frameworks:

To convert the supplied pre-trained models from TFLearn checkpoint format to protocol buffer (.pb) format (used by OpenCV DNN, TensorFlow, ...) do:

$ cd converter
$ python
$ python

This creates three .pb files inside the converter directory (firenet.pb / inceptionv1onfire.pb/sp-inceptionv1onfire.pb) which can then be tested with the OpenCV DNN module (for example, using OpenCV > 4.1.0-pre) from within the same directory:

$ python

(N.B. for the superpixel network, the test script just checks loading and inference with the .pb loaded model but does not supply an actual superpixel image - just any test image, hence inference fails to detect the fire correctly for the example only).

To convert to to other frameworks (such as PyTorch, MXNet, Keras, ...) from this tensorflow format (protocol buffer, .pb): - please see the extensive deep neural network model conversion tools offered by the MMdnn project.

Example video:

Examples Video Example - click image above to play.


If making use of this work in any way (including our pretrained models or dataset), you must reference the following article in any report, publication, presentation, software release or any other materials:

Experimentally defined Convolutional Neural Network Architecture Variants for Non-temporal Real-time Fire Detection (Dunnings and Breckon), In Proc. International Conference on Image Processing IEEE, 2018.

  author =     {Dunnings, A. and Breckon, T.P.},
  title =      {Experimentally defined Convolutional Nerual Network Architecture Variants for Non-temporal Real-time Fire Detection},
  booktitle =  {Proc. International Conference on Image Processing},
  pages =      {1558-1562},
  year =       {2018},
  month =      {September},
  publisher =  {IEEE},
  doi = 	 {10.1109/ICIP.2018.8451657},
  keywords =   {simplified CNN, deep learning, fire detection, real-time, non-temporal, non-stationary visual fire detection},

In addition the terms of the LICENSE must be adhered to.


Atharva (Art) Deshmukh (Durham University, github and data set collation for publication).

You can’t perform that action at this time.