关键词: 森林火灾, 图像识别, 火焰检测, YCbCr 颜色空间, 局部二值模式, 局部相位量化, 支持向量机

Abstract: A flame recognition algorithm based on the partitioned LBP histogram feature combined with the LPQ histogram feature was proposed according to the unique color and texture feature of flame． The algorithm was de-signed to reduce the false positive rate of forest fire in the presence of flame-like interference source and increase the speed of fire warning． Firstly，the rule in YCbCr color space was used to detect the suspected flame region．Secondly，LBP and LPQ were used to extract the texture from the spatial domain and frequency domain． Then the feature vector was obtained by combining the extracted texture features． Finally，the feature vector was inputted into support vector machine (SVM) for flame recognition． The experimental results show that the algorithm is ro-bust and has a high detection rate． When there is a flame-like interference source，the accuracy of flame iden-tification of the test set can reach 94.55%． Compared with deep learning algorithm，the proposed algorithm can signifi-cantly improve the speed of fire warning while ensuring a high accuracy． Its forecasting time is 1/4 of the forecasting time of DBN，and 1/50 of that of CNN． Thus the algorithm provides a basis for fast and accurate forest fire warning．

Key words: forest fire, image identification, flame detection, YCbCr color space, local binary pattern, local phase quantization, support vector machine