关键词: 火灾识别, 谱值特征, 最优特征选择, Gentle Boosting

Abstract: This paper presents a new method for fire detection in digital image sequences.Firstly，the suspicious fire-region is located by the intensity information of each image，and then the features constituted by intensity-variation and fire area-variation of each pixel in the suspicious fire-region are obtained.Meanwhile，a new spectrum feature is proposed to eliminate the disturbance from the regular flickering light source.Next，the categorizer is designed in Gentle Boosting algorithm，which accomplishes the work of optimal feature selection in the same time of training.With this algorithm，the characteristic space can be effectively diminished and the features can be well classified.Finally，through experiment，the recognizing results in different environments and the loca-ting results of a weak fire show that the method has a high accuracy as well as an economical computing time.

Key words: fire recognition, spectrum feature, optimal feature selection, Gentle Boosting