计算机与现代化
計算機與現代化
계산궤여현대화
COMPUTER AND MODERNIZATION
2015年
8期
121-126
,共6页
图像分类%单一编码%最小二乘SVM%弱分类器%SCOBoost
圖像分類%單一編碼%最小二乘SVM%弱分類器%SCOBoost
도상분류%단일편마%최소이승SVM%약분류기%SCOBoost
image classification%simplex coding%LS-SVM%weak classifier%SCOBoost
现有图像分类机制一般将多类别分类问题划分成多个二类别分类问题的集合进行解决,类别数的多少直接影响着二值分类器的需求量。由于图像分类问题牵扯的类别数通常较多,从而导致其训练时间过长、计算需求过高以及测试代价过大等。针对上述问题,本文设计一种新型的多分类boosting优化算法,即SCOBoost。首先,以单一编码技术为基础,结合最小二乘支持向量机( LS-SVM)目标函数,提出单一编码的多分类改进目标;其次,选取其数量与类别数无关联的弱分类器集合作为核函数,利用boosting的递归方式进行求解。通过对不同数据集实验,结果表明SCOBoost不仅拥有较高的分类性能,而且具有算法复杂度低、训练时间不受类别数影响、测试速率快等优点。
現有圖像分類機製一般將多類彆分類問題劃分成多箇二類彆分類問題的集閤進行解決,類彆數的多少直接影響著二值分類器的需求量。由于圖像分類問題牽扯的類彆數通常較多,從而導緻其訓練時間過長、計算需求過高以及測試代價過大等。針對上述問題,本文設計一種新型的多分類boosting優化算法,即SCOBoost。首先,以單一編碼技術為基礎,結閤最小二乘支持嚮量機( LS-SVM)目標函數,提齣單一編碼的多分類改進目標;其次,選取其數量與類彆數無關聯的弱分類器集閤作為覈函數,利用boosting的遞歸方式進行求解。通過對不同數據集實驗,結果錶明SCOBoost不僅擁有較高的分類性能,而且具有算法複雜度低、訓練時間不受類彆數影響、測試速率快等優點。
현유도상분류궤제일반장다유별분류문제화분성다개이유별분류문제적집합진행해결,유별수적다소직접영향착이치분류기적수구량。유우도상분류문제견차적유별수통상교다,종이도치기훈련시간과장、계산수구과고이급측시대개과대등。침대상술문제,본문설계일충신형적다분류boosting우화산법,즉SCOBoost。수선,이단일편마기술위기출,결합최소이승지지향량궤( LS-SVM)목표함수,제출단일편마적다분류개진목표;기차,선취기수량여유별수무관련적약분류기집합작위핵함수,이용boosting적체귀방식진행구해。통과대불동수거집실험,결과표명SCOBoost불부옹유교고적분류성능,이차구유산법복잡도저、훈련시간불수유별수영향、측시속솔쾌등우점。
The multi-classification problem was divided into multiple independent binary problems in existing image classification mechanism, numbers of class directly influenced demand sizes of binary classifier.The number of classes in image classification problem was very large, which led to long training time, high computing demand and high test cost.In order to effectively solve these problems, this paper designed a multi-class boosting optimizing algorithm based on simplex coding(SCOBoost).Firstly, based on simplex coding, combining with the least squares support vector machine ( LS-SVM) objective function, this paper pro-posed multi-classification improvement goal based on simplex coding;secondly, selected the weak classifiers which are not associ-ated with the number of classes as the kernel function, and used iterative methods of boosting to solve.Experiment results on dif-ferent data sets showed, SCOBoost not only had higher classification performance, but also had lower algorithm complexity, fast test speed and training time which is not affected by the number of classes and so on.