Lightweight Convolutional Neural Network for Fire Classification in Surveillance System

Abstract

Fire is one of the worst disasters for human life. Fire can happen anywhere and the leading cause can be natural or man. Over the last century, scientists have invented sensor-based methods to minimize damage and provide early warning of fires. However, these applications are only applied in a limited space and distance. For the purpose of fire remote warning and deploying on low-computing devices, this paper proposes a vision-based method using a lightweight convolutional neural network architecture combined with the inception and attention mechanisms. This proposed network includes two main modules: a feature extractor and a classifier. The feature extractor exploits convolution layers, depthwise separable convolution layers, inception module, and attention mechanism to extract high-level feature maps. Next, the classifier applies the global average pooling layer to quickly reduce the feature map dimensions and uses the softmax function to calculate the probability of each class. The experiments performed the training and evaluation on six datasets with an accuracy of over 96%. The fire surveillance system was implemented with simulation videos on GPU, CPU, and Jetson Nano devices, with the highest speeds of 200.95 FPS, 31.08 FPS, and 14.27 FPS, respectively. A set of demonstration videos, source code, and proposed dataset are provided here: https://bit.ly/3Wlpycf.