基于变形决策树的图像方向判别和应用

【作者】 朱赟；

【导师】 吴炜； 倪成凤；

【作者基本信息】 上海师范大学 ， 计算机应用技术， 2003， 硕士

【摘要】 本课题为日本欧姆龙软件股份有限公司发掘可用于图像方向分类的图像特征量并实现一个判别图像方向（正立方向和两种非正立方向）的分类系统。课题的研究关注基于内容图像检索分类研究中有待解决的主要问题：底层图像特征与高层图像概念的结合。本课题使用图像轮廓线向量特征来反映图像的构图特征，通过教师实例获得用户的方向分类概念，然后利用决策树的学习建立分类模型，并且在此分类模型的基础上实现了一个图像正立方向判别系统。判别系统通过在较高抽象层次上的对象描述技术实现了图像特征量的易扩充性、细节信息的高隐蔽性以及较高的系统安全性。实验结果表明，系统所使用的轮廓线向量图像特征也能够较有效地应用于图像方向分类，而机器学习则能够有效地为之建立决策树分类模型。在现有分类系统基础上，本文于提高系统性能方面做了一些改进和应用。一方面是提高系统的学习和分类效率：在图像数量庞大和分类种类繁多的情况下，决策树的建立和探索效率直接影响到系统性能，我们利用变形决策树来提高分类模型建立和探索的效率，以及决策树局部学习来提高增量式学习的效率。另一方面是提高教师实例选择的效率。本文将加强分类边缘和确认不确定分类的未标识实例两方面结合，实现教师实例自动选择。实验结果说明使用变形决策树和教师数据自动选择的方法后，系统性能得到了一定程度的提高。更多还原

【Abstract】 This thesis introduces a contract research for Omron Software Co. Ltd Japan. The research includes development of image characteristics describing image direction and then constructing an image direction classification system. In the area of contents based image classification, building the relationship between the low-leveled image characteristics and the high-leveled contents based concept is one of the most important works. In our research, variation of contour line is used for describing the characteristics of image structure. Using teacher images and machine learning method, an image direction classification model is built as a decision tree. Test results argued the validity of this method. Based on such model, we developed an image direction detecting system. The system also uses an object-oriented technique on highly abstract level, which implements easily extension of image characteristics and highly hidden of characteristic details, so as to ensure a relatively high system stability.Efforts in improving the capability of model are made. While the number of identified class increases, decision trees will become very large and deep. We deformed the decision tree to further improve the balance and the constructing efficiency of the decision tree. On the other hand, a method for auto-selecting of teacher images is implemented. The mechanism tries to select images near the boundary between different classes as well as the most uncertainty samples. The test results show the improvements in system capability.更多还原