Smoke Detection from Different Environmental Conditions Using Faster R-CNN Approach Based on Deep Neural Network

Abstract

From the last few decades, smoke detection performed for noble purposes like to rescue people from fire, make wood-land or wildlife safe from fire disaster and so on. Most of those detections were sensor based where detectors detect smoke optically or by physical processes and which causes false alarm most of the time. By the passing time, the author’s tries to overcome those false alarm rates by introducing hand-featured methods. From this perspective, those established systems performed better than sensor based tools. However, coming towards a significant point, in most instances, only one or two certain areas like forest were considered in addressing smoke. Now, moving on this research, we aimed to experience indeed with detecting diverse circumstances smoke by the Faster R-CNN approach based on the Inception-V2 deep neural network. We focused on the single class, i.e., smoke and training the method with images of our own combined extracted image frames. The proposed method achieves 97.31% detection accuracy and is compared to previous approaches to show higher detection accuracy over recent works.