【作者】 徐辉；

【导师】 王忠芝；

【作者基本信息】 北京林业大学 ， 设计艺术学， 2011， 硕士

【摘要】 在设计行业应用图像分类技术,可以使设计数字化、可管理化,让用户能便利地、准确地、高效地借鉴符合某类规定的设计作品。因此提出了一种基于表现技法的环境艺术效果图自动分类方法,目的在于利用机器学习的方法,自动建立分类的模型,以提供设计师概念化的参考借鉴浏览方式。首先,本文介绍和分析了环境设计中常用的表现技法。常用的环境设计表现技法有黑自表现、彩铅表现、水粉水彩表现、马克笔表现、计算机表现等几类。根据环境艺术设计师对不同的环境设计表现技法的感知,列出每一种表现技法所表现的低级特征,通过分析不同表现技法的视觉特点,从现有的几种特征提取算法中选择了适合艺术图像特征的算法,包括颜色、纹理等几个方面。其次,本文研究了几种重要的图像分类方法,对不同应用背景下可选用的方法进行讨论与分析,由于支持向量机(Support Vector Machine, S VM)具有坚实的理论基础与良好的分类性能等优势,本文着重研究了SVM的方法。SVM本身是一个两类问题的判别方法,不能直接应用于多类问题,从而引入了基于二叉树的多类分类算法,通过分析环境设计表达的视觉特征及SVM的原理和模型,设计了基于二叉树支持向量机(Binary Tree SVM, BT-SVM)的分类器进行环境艺术效果图的学习、训练和分类。然后从核函数参数寻优、数据归一化、样本数目三个方面进行系统评价,并对这些因素进行了归纳和总结,选择合适的分类器模型。在系统界面设计方面,提出了在MATLAB GUI中采用WebBrowser控件的方法实现图像分类的结果展示,不但实现了在MATLAB GUI窗体下浏览大量图像,同时利用HTML语言有较好的可扩展特性,实现在MATLAB用户界面中点击小图弹出大图的效果。经过实验证明,本文提出的基于SVM的环境艺术图像分类方法是行之有效的,对于采用不同表现技法创作的其他艺术作品也具有良好的分类效果。最后,本文从自适应的选择核函数、逐步加入相关反馈机制等几个方面对基于艺术风格的图像的分类技术进行了展望。更多还原

【Abstract】 The application of image classification technology in design industry enables digitization and management of design, and users can draw upon designs that conform to certain standards in a convenient, accurate and highly efficient way. The author hereby puts forward a method of automatic classification for renderings of environmental art based on expression techniques, with the aim of automatically constructing classification models by employing machine learning in order to provide designers with a conceptualized browsing method for reference.First, this paper offers an introduction to and analysis of common expression techniques that are utilized in environmental design, including black-white, colored pencil, gouache&watercolor, marker, computer expression, etc. The low-level features reflected by each kind of expression technique are listed under the perception of environmental art designers. Through the analysis of visual characteristics of various expression techniques, this paper extracts the features of various expression techniques based on the visual differences in color, texture, smoothness and degree of saturation, and select some algorithms suitable for art images’feature extraction from several ones.Second, this paper studies several important image classification methods and discusses and analyzes the possible options under different application backgrounds. Owing to the advantages of SVM such as firm theoretical basis and excellent classification function, this paper focuses on the research of SVM. The SVM is a method for solving two-class classification problems and cannot be applied to multi-class classification problems directly; hence this paper introduces the multi-class classification algorithm based on binary tree. Through the analysis of its principles and classification methods, and the visual features of expression of environmental design, constructs a BT-SVM classifier to learn, train and classify the renderings of environmental art. This paper conducts a study on three essential factors that affect the classification effect, including parameter optimization of kernel function, data normalization and number of samples. It summarizes the advantages and shortages of these factors in order to select the most appropriate model of SVM. In respect to the design of system interface, this paper puts forward the method of adopting WebBrowser control in MATLAB GUI to realize the display of image classification results, which not only can browse many images under the MATLAB GUI window, but also realizes the effect of clicking a small image and displaying a large image in MATLAB user interface by utilizing the good scalability of HTML language. Experiments have proved that the classification of environmental art images based on SVM suggested in this paper is workable, and other art works created by utilizing different expression techniques, which also has good classification effect.Finally, this paper looks into the future of the classification technology of images based on artistic style from selection of adaptive kernel function, gradual integration of relevant feedback mechanism and other aspects.更多还原