Comparative analysis of simple rules for flame recognition
Compared to conventional fire detection techniques, high-precision computer vision-based fire detection systems have a number of desirable characteristics, such as the ability to monitor large areas, provide a bountiful amount of information, and are easy to maintain. This paper extensively and systematically investigates the use of simple color-based rules for pixel-wise flame recognition in still images. The rules are evaluated on a hundred and nineteen test images that correspond to fires in urban environments. The performances of the considered flame recognition rules are reported in terms of Recall, Balanced accuracy, Accuracy, F1 score, and Matthews correlation coefficient. The best-performing rule is identified. More complex classifiers are formed by combining two or more simple rules. The experimental results show that simple color-based rules and some of their combinations can offer effective fire recognition performance.