A novel forest fire detection system using fuzzy entropy optimized thresholding and STN-Based CNN

Abstract

In this paper, a video processing based forest fire and smoke detection using a Fuzzy Entropy optimized thresholding and convolutional neural network (CNN) based model is proposed. In our proposed architecture adopts and introduces a spatial transformer network (STN) in the CNN layer and entropy function thresholding operation in the softmax layer due to addition in the softmax layer it works excellently and accurately. We have taken initial preprocessing video quality as an essential issue, which helps us in providing less false alarm rate, we checked the blockiness, blurriness, noisiness and used efficient ways to enrich the quality of the video. After preprocessing, we propose the Three Frame Difference Method to calculate the motion of all frames. During feature extraction, we suggest an STN-Based CNN extract the dynamic features of an image. After extraction, we need to perform the classification step, and here we introduce the entropy function thresholding method in the softmax layer. The results obtained with our proposed architecture were compared with the existing methods in the literature and showed a higher exact fire detection rate, a very less false detection rate, good accuracy, and less execution time in terms of complexity. The proposed method can be used efficiently for real-time forest fire detection with a moving camera.