Video flame detection algorithm based on region growing

This paper proposes a region growing based video flame detection algorithm. Firstly, it estimates class-conditional probability density of flame and background with hand-labeled samples, and five discrimination models are proposed using maximum a-posteriori theory. Secondly, it proposes four rules for flame detection with difference of RGB channels, and ROC analysis is used to estimate rule parameters. Finally, it combines the detection results of these models and rules to detect the candidate flame regions. Region growing uses the high belief region as seed points, and some middle belief regions are classified as flame region if they are adjacent to high belief region, while other regions are classified as background regions. Experiments show that this method can achieve desired flame region in various scenes with high true positive rate and low false detection rate.