FireNet-Tiny: Very-Low Parameter Count High Performance Fire Detection Model

Abstract

In daily life, fire threats result in significant costs on the ecological, social, and economic levels. It is essential to outfit the assets with fire prevention systems due to the sharp rise in the frequency of fire mishaps. To prevent such mishaps, several studies have been conducted to develop optimal and potent fire detection models. While the earliest methods were thermal/chemical in nature, image processing was later applied for identification of fire. More recent methods have taken advantage of the significant advancements in deep learning models for computer vision. However, in order to maintain a suitable inference time (leading towards real-time detection) and parameter count, the majority of deep learning models have to make trade-offs between their detection speed and detection performance (accuracy/recall/precision). The very lightweight convolution neural network we offer in this paper is specifically designed for the fire detection use case. The proposed model can be embedded in real-time fire monitoring equipment and could also prove potentially useful for future fire monitoring methods such as unmanned aerial vehicles (drones). By further diminishing the trainable parameter count of the model, the fire detection results obtained using the proposed FireNet-Tiny significantly outperform the prior low parameter count models. When tested against FireNet dataset, FireNet-Tiny, which only comprises 261,922 parameters, was shown to have an overall accuracy of 95.75%. In comparison, FireNet-v2 provided 94.95% accuracy with 318,460 parameters.