Fire detection in video surveillance using superpixel-based region proposal and ESE-ShuffleNet

This paper proposes a forest fire detection framework using superpixel-based suspicious fire region proposal and light-weight convolutional neural network. The proposed methodology contains two main steps. In suspicious fire region proposal, we introduce a novel superpixel algorithm (SCMM) driven by Cauchy mixture model. Then, the negative Under-segmentation Error (UE) of each superpixel is applied to inter-frame comparison for predicting varying superpixels. After that, by computing the features of motion superpixels using Local Difference Binary (LDB) descriptor for two adjacent frames, the suspicious fire regions are localized. In following fire identification, to improve network performance while reducing computational complexity, this study presents a light-weight network architecture, called Expanded Squeeze-and-Excitation ShuffleNet (ESE-ShuffleNet). All suspicious fire regions are sent into this network to identify as either fire or non-fire included. Experiments show that our framework performs well on fire detection tasks. Code is available at http://www.imagetech-polynomials.com/ESE-ShuffleNet.html.