交互式训练样本获取方法  

Interactive obtaining method for training samples

作　　者：陈蔼祥[1] 

机构地区：[1]广东财经大学数学与统计学院,广州510320

出　　处：《计算机应用》2015年第A01期278-282,共5页journal of Computer Applications

基　　金：国家自然科学基金国际合作与交流项目(中德合作)(61111130183);教育部重点项目(210257);广东省自然科学基金资助项目(10451032001006140);广州市科技和信息化局应用基础研究项目(10C12140131);广东省教育厅普通高校育苗工程项目(LYM10081)

摘　　要：有监督学习算法是机器学习中的一类重要算法,该类算法要求外界提供含监督信号的样本作为训练数据。虽然机器学习领域提供了许多基准测试数据,但很多情况下需要自己生成训练样本。给出了一种交互式训练样本获取方法:通过对原始图像进行一种或多种混合的随机变换,用户挑选那些能被人眼识别的样本作为有效样本加以保存。实验结果表明,所提方法产生的图片能模拟摄像头在不同角度、姿态、光照、遮挡等各种复杂场景下拍摄的图像的效果。用系统生成的训练样本训练朴素贝叶斯(NB)分类器,能达到95.042%的识别精度,结果优于UCI人工字符集训练同样的NB分类器时88.487 5%的识别精度。As one type of machine learning, supervised learning algorithm needs labeled samples as its training data. In many cases, one need to create labeled samples on his own, although there already exist many benchmark data in machine learning field. To address the issuse on how to construct training data for most exist classifiers, a system was developed to help user to pick the pictures what they want. The system loaded a seed picture, then sequentially performed one or several random transforms on the seed picture, and the transform results would be shown in a window such that user could pick the pictures which could be discriminated by human eyes. Experiment results show that the system can easily generate satisfactory samples, use these samples as training samples of Naive Bayes （ NB） classifier, and the accuracy on test set can reach 95. 042%, better than the 88. 487 5% of the same classifier on UCI artificial letter data set.

关 键 词：有监督学习 训练样本 仿射变换 分类器 朴素贝叶斯 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]