A Study of Rule-based Flame Pixel Classifiers in Still Images

Abstract

Southern Europe is an exceptionally fire-prone region. Effective prevention, early warning, and early response are important actions that are needed to reduce disastrous consequences on people and structures. Flame pixel classification is the first and often critically important stage of fire detection systems based on computer vision. In this paper, a number of rule-based flame pixel classifiers in still images are studied based on the BoWFire fire dataset. The performances of the considered rule-based flame pixel classifiers are reported in terms of F1 score, Matthews correlation coefficient, and Balanced accuracy. The experimental results demonstrate the effectiveness of simple rule-based flame pixel classifiers for flame detection in still images.