Developing an Integrated Computer Vision Image Processing Technique to raise an Early Alarm for Fire Detection

Abstract

In Computer Vision and image classification task, convolutional neural networks (CNNs) have demonstrated high performance. Their use in fire detection systems will make detection much more accurate, reducing the number of fire disasters and their ecological and social effects. However, implementation in real-world surveillance networks of CNN-based fire detection systems poses the greatest risk due to their high inference memory and computational requirements. An original, energy-efficient, and computationally efficient design is presented in this paper. Squeeze Net-inspired CNN architecture for fire detection, localization, and semantic comprehension of the fire scene. The experimental results show that our proposed solution achieves accuracies comparable to those of other more complex models, primarily due to its increased depth, despite its low computational requirements. Because it doesn't have any dense, fully connected layers and uses smaller convolutional kernels, its computational requirements are reduced. This paper also shows how fire detection efficiency and accuracy can be traded off by taking into account the particulars of the problem at hand and the variety of fire data.