虛警率約束的聯(lián)合弱分類(lèi)器集成學(xué)習(xí)算法
發(fā)布時(shí)間:2018-12-14 16:15
【摘要】:提出一種聯(lián)合弱分類(lèi)器集成學(xué)習(xí)算法。借鑒Adaboost方法采用弱分類(lèi)器構(gòu)建強(qiáng)分類(lèi)器的思想,聯(lián)合多個(gè)弱分類(lèi)器構(gòu)建特征分類(lèi)的得分函數(shù),生成一個(gè)集成分類(lèi)器。在分類(lèi)器訓(xùn)練時(shí),采用ROC曲線圍成的AUC面積值構(gòu)建目標(biāo)函數(shù),加入虛警率上下限約束條件,采用列生成算法學(xué)習(xí)弱分類(lèi)器,采用割平面法學(xué)習(xí)弱分類(lèi)器的系數(shù)。在PASCAL VOC-2007數(shù)據(jù)集上進(jìn)行目標(biāo)檢測(cè)實(shí)驗(yàn),實(shí)驗(yàn)結(jié)果表明,與常用的支持向量機(jī)、Adaboost、隨機(jī)森林和卷積神經(jīng)網(wǎng)絡(luò)分類(lèi)方法相比,該方法的假正率指標(biāo)低,真正率指標(biāo)高。
[Abstract]:A joint weak classifier ensemble learning algorithm is proposed. Using the idea of weak classifier to construct strong classifier using Adaboost method and combining several weak classifiers to construct the score function of feature classification, an integrated classifier is generated. In classifier training, the objective function is constructed by the AUC area value surrounded by ROC curve, the upper and lower bound constraints of false alarm rate are added, the weak classifier is learned by column generation algorithm, and the coefficient of weak classifier is studied by cutting plane method. The experimental results of target detection on PASCAL VOC-2007 data sets show that the false positive rate index is lower and the real rate index is higher than the usual classification methods such as support vector machine, Adaboost, stochastic forest and convolution neural network.
【作者單位】: 蘇州信息職業(yè)技術(shù)學(xué)院計(jì)算機(jī)科學(xué)與技術(shù)系;鄭州財(cái)經(jīng)學(xué)院信息工程學(xué)院;太原理工大學(xué)電氣工程學(xué)院;
【基金】:河南省科技廳科技計(jì)劃課題基金項(xiàng)目(112102310550)
【分類(lèi)號(hào)】:TP181
本文編號(hào):2378929
[Abstract]:A joint weak classifier ensemble learning algorithm is proposed. Using the idea of weak classifier to construct strong classifier using Adaboost method and combining several weak classifiers to construct the score function of feature classification, an integrated classifier is generated. In classifier training, the objective function is constructed by the AUC area value surrounded by ROC curve, the upper and lower bound constraints of false alarm rate are added, the weak classifier is learned by column generation algorithm, and the coefficient of weak classifier is studied by cutting plane method. The experimental results of target detection on PASCAL VOC-2007 data sets show that the false positive rate index is lower and the real rate index is higher than the usual classification methods such as support vector machine, Adaboost, stochastic forest and convolution neural network.
【作者單位】: 蘇州信息職業(yè)技術(shù)學(xué)院計(jì)算機(jī)科學(xué)與技術(shù)系;鄭州財(cái)經(jīng)學(xué)院信息工程學(xué)院;太原理工大學(xué)電氣工程學(xué)院;
【基金】:河南省科技廳科技計(jì)劃課題基金項(xiàng)目(112102310550)
【分類(lèi)號(hào)】:TP181
【相似文獻(xiàn)】
相關(guān)期刊論文 前3條
1 郭立娟;;非線性弱分類(lèi)器的存在性[J];蘭州理工大學(xué)學(xué)報(bào);2012年02期
2 謝紅躍;方昱春;蔡起運(yùn);;一種新的改進(jìn)AdaBooat弱分類(lèi)器訓(xùn)練算法[J];中國(guó)圖象圖形學(xué)報(bào);2009年11期
3 趙梅芳;羅阿理;吳福朝;趙永恒;;自適應(yīng)增強(qiáng)方法在光譜自動(dòng)分類(lèi)中的應(yīng)用[J];光譜學(xué)與光譜分析;2008年02期
,本文編號(hào):2378929
本文鏈接:http://www.sikaile.net/kejilunwen/zidonghuakongzhilunwen/2378929.html
最近更新
教材專著