Codes Used on My Thesis Work - Supervised Methods on Fire and Smoke detection ( FireFront Project ). All the codes produced are ment to be executed on Google Colab environment using Python3.
On first part of the project I worked on a global image classifier using different networks. The objective was to detect if a certain image contains fire or not, not giving any information about the localization of that fire.
Networks used:
AlexNet: code
SqueezeNet: code
Then I moved on to segmentation networks in which I used fully supervised training methods to train. The networks are able to segment the given images in areas of fire and non-fire ( binarization )
Networks used:
U-Net: code
Dataset created still to be published...
In order to solve the multi-scale problem of detection we use a Quad-Tree algorithm to dynamically slice the input images into smaller patches to help the network detect smaller regions of the phenomenon. The next code file includes the Quad-Tree implementation and the calculation for the system performance:
Global System: code
The saved parameters for both networks, for both classes, can be downloaded below:
SqueezeNet Fire : code
SqueezeNet Smoke : code
U-Net Fire: code
U-Net Smoke: code