考虑颜色特征最优组合的CART决策树火灾图像识别方法*

针对火灾图像识别过程中颜色特征数量多、特征间相关性复杂、难以在多维特征融合过程中有效融合图像颜色特征等问题，提出1种考虑颜色特征最优组合的CART决策树火灾图像识别方法。首先，在Lab、RGB、HSV 3种色彩模式下基于图像颜色特征提取火灾图像特征序列；其次，分别在3种色彩模式下基于精细决策树与特征随机排列组合方法提取颜色特征中最优组合特征；最后，将提取的火灾图像最优组合特征序列作为CART决策树输入进行模型训练，并通过测试样本以及其他机器学习方法进行模型泛化能力的分析。研究结果表明:本文方法寻找出识别火灾图像的最优颜色特征组合为“Kb1+Var1+Kg+Kb2+Var2+Kh+Ks+Kv”；CART决策树方法对于火灾图像识别的测试准确度可达84.5%，其分类效果明显优于其他决策树类与集成树类方法；9折为最佳交叉验证折数，其测试准确度可达86.47%，与5折交叉验证相比明显提升14.77%。研究结果可为火灾图像识别提供方法基础。

Fire image recognition method of CART decision tree considering optimal combination of color features

Aiming at the problems of large number of color features,complex correlation between features,and difficulty in effectively integrating the image color features in the process of multi-dimensional feature fusion in the process of fire image recognition,a fire image recognition method of classification and regression tree (CART) decision tree considering the optimal combination of color features was proposed.Firstly,the fire image feature sequence was extracted based on the image color features in three color modes of Lab,RGB and HSV.Secondly,the optimal combination features were extracted from the color features based on the fine decision tree and feature random permutation and combination method in three color modes,respectively.Finally,the extracted optimal combination feature sequence of fire image was used as the input of the CART decision tree to train the model,and the generalization ability of the model was analyzed through test samples and other machine learning methods.The results showed that the optimal color feature combination for recognizing the fire images found by this method was “Kb1+Var1+Kg+Kb2+Var2+Kh+Ks+Kv”.The test accuracy of fire image recognition by CART decision tree method could reach 84.5%,and its classification effect was significantly better than other decision tree and ensemble tree methods.9 fold was the best cross-validation fold number,and its test accuracy could reach 86.47%,which was significantly improved by 14.77% compared with 5 fold cross-validation.The research results can provide a method basis for the fire image recognition.