基于码字匹配和引力筛选的半监督协同训练算法  

Semi-Supervised Collaboration Training Algorithm Based on Codeword Matching and Gravitation Selecting

作　　者：陈善学[1] 尹修玄 杨亚娟[1] 

机构地区：[1]重庆邮电大学移动通信技术重庆市重点实验室,重庆400065

出　　处：《武汉大学学报（信息科学版）》2015年第10期1386-1391,1408,共7页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金资助项目(61271260;61102062);重庆市教委科学技术研究资助项目(KJ1400416)~~

摘　　要：针对传统的Co-training和Tri-training协同训练算法中基分类器独立性低、迭代过程中误差累积和整体泛化性能低的问题,将多视图理论、编码理论和万有引力公式引入协同训练分类算法中,提出了改进算法,算法有效地防止了迭代过程中的误差累积,同时提高了分类系统的泛化性能。在高光谱图像分类实验中,随机地从数据集中抽取5%、10%和20%样本作为已标记训练集时,码字匹配的协同训练分类算法对比Co-training和Tri-training算法,在分类精度上平均分别提高了12.38%和6.13%,在Kappa系数上平均分别提高了0.2和0.07。进一步加入引力筛选机制,对比Co-training和Tri-training算法,在分类精度上平均分别提高了21.30%和10.99%,在Kappa系数上平均分别提高了0.26和0.13,结果表明了本文算法的有效性。Traditional collaboration training algorithms,such as Co-training and Tri-training,has some problems：low independence of the base classifier,error accumulation during iteration and low generalization system performance.Accordingly,this paper proposes an import multi-view theory,coding theory and formula for universal gravitation as applied to the collaboration training algorithm preventing both error accumulation and improving generalization performance at the same time.During a hyper-spectral image classification experiment,randomly selecting 5%,10% and 20% samples from data sets as a labled train set,the collaboration training algorithm for codeword matching had a12.38% and 6.13% higher accuracy matching rate than Co-training and Tri-training respectively.At the same time,it had a respective 0.2and 0.07 higher Kappa coefficient.In contrast,a collaboration training algorithm for classification based on codeword matching and gravitation selecting had 21.30%and 10.99% higher accuracy than Co-training and Tri-training and 0.26 and 0.17 higher Kappa coefficient.These results demonstrate the validity of the proposed algorithm.

关 键 词：高光谱图像分类 协同训练 码字匹配 万有引力 

分 类 号：P237.3[天文地球—摄影测量与遥感]