FireNet-Micro: Compact Fire Detection Model with High Recall

Abstract

Fire occurrences and threats in everyday life incur substantial costs on ecological, economic, and even social levels. It is crucial to equip establishments with fire prevention systems due to the notable increase in fire incidents. Numerous studies have been conducted to develop efficient and optimal fire detection models in order to prevent such mishaps. Initially, thermal/chemical methods were used, but later, image processing techniques were also employed to identify fire occurrences. Recent approaches have capitalized on the advancements in deep learning models for computer vision. However, most deep learning models face a trade-off between detection speed and performance (accuracy/recall/precision) to maintain a reasonable inference time (for real-time applications) and parameter count. In this paper, we present a bespoke and highly lightweight convolutional neural network specifically designed for fire detection. This model can be integrated into real-time fire monitoring equipment and potentially applied in future methods suhc as CCTV surveillance cameras, traffic lights, and unmanned aerial vehicles (drones) for fire monitoring in futuristic smart city scenarios. Despite having significantly fewer trainable parameters, our customized model, FireNet-Micro, outperforms existing low-parameter-count models in fire detection. When evaluated on the FireNet dataset, FireNet-Micro, with only 171,234 parameters, achieved an impressive overall accuracy of 96.78%. In comparison, FireNet-v2 attained 94.95% accuracy with 318,460 parameters (which is almost double the parameter count of the proposed FireNet-Micro).