Fire/Flame Detection with Attention-Based Deep Semantic Segmentation

Abstract

Fire/flame detection from images or videos is very important for early fire warning systems. In this way, fires can be intervened early and extinguished before they grow. Recently, many studies have been published on early fire warning systems based on image processing and machine learning. These studies are generally color space-based image segmentation applications. The given images are first transferred to another color space, and the fire/flame regions are determined by using color segmentation. In this study, a segmentation technique using deep network architecture for fire/flame detection is presented. The proposed method is a segmentation network structure in which the attention gate module is integrated. In the presented method, the success of the deep network architecture is evaluated by using the dice, Tversky, and focal Tversky loss functions. A data set containing 500 images was used for experimental studies, with the fivefold cross-validation criterion, and the success achieved was presented depending on the mean dice and Jaccard similarity criteria. The calculated results were compared with some studies in the literature. The comparison results were shown that the presented technique produced more successful results.