Tested using Python 3.4.6, TensorFlow 1.9.0, and OpenCV 3.3.1

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

                                                                            FireNet architecture (above)                                                                           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]

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 firenet.py and inceptionV1-OnFire.py files corresponding to the two binary (full-frame) detection models from the paper. In addition the superpixel-inceptionV1-OnFire.py 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 download-models.sh which will create an additional models directory containing the network weight data.

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.

Training datasets:

• The custom dataset used for training and evaluation can be found on Durham Collections (together with the trained network models).
• In addition, standard datasets such as furg-fire-dataset were also used for training and evaluation.

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

$ git clone https://github.com/tobybreckon/fire-detection-cnn.git
$ cd fire-detection-cnn
$ sh ./download-models.sh
$ python firenet.py models/test.mp4
$ python inceptionV1-OnFire.py models/test.mp4
$ python superpixel-inceptionV1-OnFire.py models/test.mp4

Video Example - click image above to play.

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.

@InProceedings{dunnings18fire,
  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 =      {1-5},
  year =       {2018},
  month =      {September},
  publisher =  {IEEE},
  keywords =   {simplified CNN, fire detection, real-time, non-temporal, non-stationary visual fire detection},
}

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