一种基于多尺度局部纹理特征和CART决策树的野外火灾火焰图像识别算法  

A FIELD FIRE FLAME IMAGE RECOGNITION ALGORITHM BASED ON MULTI-SCALE LOCAL TEXTURE FEATURES AND CART DECISION TREE

作　　者：冯丽琦 赵亚琴[1] 孙一超 龚云荷 Feng Liqi;Zhao Yaqin;Sun Yichao;Gong Yunhe(School of Mechanical and Electronic Engineering,Nanjing Forestry University,Nanjing 210037,Jiangsu,China)

机构地区：[1]南京林业大学机械电子工程学院,江苏南京210037

出　　处：《计算机应用与软件》2019年第5期194-198,共5页Computer Applications and Software

基　　金：国家自然科学基金青年科学基金项目(31200496);江苏省高等学校大学生创新创业训练计划项目SPITP(201710298011Z)

摘　　要：为了消除野外环境中枯草、枯树枝、枯树叶等干扰对象对野外火灾识别的影响,提高火焰识别的准确率,提出一种新的基于Gabor滤波和局部二值模式(LBP)的多尺度局部纹理特征提取方法,并构建Adaboost-SVM火焰图像分类器。利用火焰的颜色特征提取出疑似火焰区域;对疑似火焰区域进行Gabor滤波,再对Gabor滤波后不同尺度下的图像以16×16的像素邻域网格作为采样窗口,采用LBP提取其纹理特征;运用CART决策树对LBP特征向量进行降维,将分类回归树算法(CART)选择出来的特征输入到支持向量机(SVM)训练分类器,进行火灾火焰图像识别。实验结果表明,野外火灾火焰的识别准确率为96%,证明了该算法的有效性。In order to eliminate the influence of the interference objects such as withered grass,withered branches and dead leaves on the fire recognition in the wild environment and improve the accuracy of the fire recognition,we presented a new multi-scale local texture feature extraction method based on Gabor filter and local binary pattern(LBP)and constructed an Adaboost-SVM flame image classifier.Color features of flame were used to extract the suspected flame regions,on which the Gabor filtering was performed.For the images at different scales after filtering,the 16×16 pixel neighborhood grid was used as the sampling window,and the texture features were extracted by LBP.We adopted CART decision tree to reduce the dimensionality of LBP eigenvector,and input the features selected by CART into the SVM training classifier for fire flame image recognition.Experimental results show that the correct rate of the fire flame recognition reaches 96%,which proves the effectiveness of the algorithm.

关 键 词：火焰识别 UNIFORM LBP GABOR滤波 CART决策树 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]