CAJ | 학술논문

图像型火灾探测实际上是不平衡数据的二分类问题，现有方法在处理不平衡数据分类问题时常常会引入新的噪声点或丢掉很重要的信息，算法稳定性较差。根据Adaboost对样本分配不同权重的优势，和SVM在平衡数据条件下较好的分类性能，将Adaboost算法和支持向量机（SVM）相结合，提出Adaboost-SVM算法。把火焰疑似区域的特征值作为SVM分类器的输入参数，利用Adaboost算法重点标记错分样本，并对样本的权重设定阈值，采用一定的准则对少数样本进行再构造使正负样本达到平衡。最后在训练数据的同时，通过投票机制输出最终分类结果。实验结果表明，该算法提高了火灾在正负样本分布不平衡时的分类性能。
도상형화재탐측실제상시불평형수거적이분류문제，현유방법재처리불평형수거분류문제시상상회인입신적조성점혹주도흔중요적신식，산법은정성교차。근거Adaboost대양본분배불동권중적우세，화SVM재평형수거조건하교호적분류성능，장Adaboost산법화지지향량궤（SVM）상결합，제출Adaboost-SVM산법。파화염의사구역적특정치작위SVM분류기적수입삼수，이용Adaboost산법중점표기착분양본，병대양본적권중설정역치，채용일정적준칙대소수양본진행재구조사정부양본체도평형。최후재훈련수거적동시，통과투표궤제수출최종분류결과。실험결과표명，해산법제고료화재재정부양본분포불평형시적분류성능。
Image fire detection is actually a binary classification problem of imbalanced data,existing technology often brings in new noise points or loses very important messages when dealing with the imbalanced data classification problems,and the stability in algorithm is also poor.Accord-ing to the advantage of Adaboost in assigning different weights to samples,plus the SVM has a better classification performance under equilibrium data condition,we propose the Adaboost-SVM algorithm by combining Adaboost algorithm with support vector machine (SVM).It takes the eigen-value of suspected flame area as the input parameter of SVM classifier,intensively marks the fault samples with Adaboost algorithm,and sets the threshold value on sample weights,and reconstructs a few samples by adopting a certain criteria to reach the balance between positive and negative samples.At last,the algorithm outputs final classification results through voting mechanism when training data.Experimental results show that this algorithm improves the classification performance of fire when the distribution is imbalance between positive and negative samples.