Low-complexity high-performance deep learning model for real time low cost embedded fire detection system.

Abstract

Correct and timely detection of fires has been an active area of research. Both shallow learning (with manual feature engineering), as well as deep learning (with its promise of automatically extracting meaningful representations from the data) approaches, have been used to solve fire detection problems. Most deep learning systems outperform the hand-crafted algorithms for fire detection, particularly due to the enormous potential offered by Convolutional Neural Network and its variants. The design problem is further compounded when the model is intended to be deployed on a low-computationally-intensive portable and mobile hardware. This requirement calls for a model which has a suitably small size on disk (translating to a lesser number of parameters to be estimated). Although some MobileNet based solutions are available which are superior to their counterparts (both in terms of increased accuracy as well as reduced complexity), there is still scope for improvement. The present work endeavors to demonstrate this by proposing a modified MobileNetV2 architecture and a better transparent data handling strategy that is capable of outperforming the existing solutions while being computationally viable for deployment on less able hardware.