Early Fire Detection and Alert System using Modified Inception-v3 under Deep Learning Framework
Anuraag Biswas, Swarup Kr Ghosh, Anupam Ghosh
Abstract
ADSI-2019 report states that there were approximately eleven thousand fire accidents reported across India in 2019. In this article, a deep learning model has been developed for fire and smoke detection from images. In this context, we have modified the Inception Convolution Neural Network of the third version (known as Inception-V3) model which is commonly applied for image classification and object discovery tasks for this work. The modified Inception-V3 has been applied on the fire image dataset including smoke for fire image detection. A new optimization function has been incorporated into this model which reduces the computational cost in an efficient way. We have compared the results such as deep learning-based models as well as early fire detection. In this study, the Inception v3 based model yielded the best result with the least number of false positives as compared to the early investigations.