Method design of small-scale fire detection

In this paper, we propose a small-scale fire detection method. This fire detection method consists of two steps. The first step of the proposed method exploits the color and the intensity variation information of the pixels to determine the candidate to the fire region. The second step refines the results of the first step to obtain the ultimate fire region. The first and second steps have the following merits and effects: the first step is able to detect almost all of the true fire regions but it also takes some non-fire regions as the fire, which are referred to as 'false' fire regions. On the other hand, the second step can greatly eliminate the 'false' fire regions generated from the first step. The second step achieves this by using a learning method that first captures a number of training samples of the moment features of the true and false fire regions generated from the first step and then exploits the training samples and an improved K nearest neighbor (KNN) classifier to produce the ultimate detection result. The experimental results show that the proposed method performs very well in detecting small-scale fire.