【作者】 周长春；

【导师】 孙建伶；

【作者基本信息】 浙江大学 ， 计算机技术， 2015， 硕士

【摘要】 随着图像处理技术和互联网技技术的发展进步,越来越多的实体美术作品被数字化并展示于互联网之上。如何更加有效地组织并管理以使得用户能更加简便快捷访问这些数字化的美术作品成为了当前急需解决的问题。本文系统综述了基于内容的图像检索技术,包括特征提取、支持相似查询的高维数据索引技术,以及相关的图像分类技术,包括图像的稀疏编码技术、分类器模型。考虑美术作品的自身特点,综合现有的图像检索技术,本文实现了一个美术作品检索系统。艺术风格作为美术作品的视觉特征之一,它表征的是图片的高阶语义,传统的以风格来对作品进行分类主要是通过人为的手段来进行的,是基于人的相关知识和经验。由于美术作品数量巨大,使得仅仅通过手工的方式来进行分类变得不可行。为此本文在图像分类及相关技术基础之上,提出了两种基于SIFT特征的美术作品风格分类方案：基于主题模型的风格分类以及基于金字塔匹配的风格分类方案。基于金字塔匹配的分类方案其主要思路：1)抽取美术作品的SIFT特征,用以表征作品的风格信息；2)对抽取出的特征进行稀疏编码,以更加简洁有效地表示图片的风格信息。3)使用金字塔核函数来度量美术作品在风格上的相似性：4)训练一个线性的Multi-class SVM模型来预测美术作品所属风格。与基于金字塔核函数的分类方案相比,基于主题模型的分类方案同样是建立在SIFT特征的词袋表示之上,不同之处在于两者对特征集合的相似性度量。主题模型分类方案以图像的词袋表示作为输入,训练主题模型来得到图像所对应的主题分布,利用主题分布向量间的欧式距离来度量美术作品间的相似性,而最终分类器的输入则相应的变为图像的主题分布。更多还原

【Abstract】 Along with the development of image processing and network technology, more and more art works were digitalized and exhibited on the internet. How to effectively organize and manage the works so that people can browse and retrieve them more easily has become an imperative problem need to be addressed.This thesis summarized content based image retrieval and image classification technology, such as visual feature extraction, metric space based high-dimensional data index technology, sparse coding, image classifier. With regard to the characters of art works, this thesis proposed and designed an image retrieval system based on text and content image retrieval technology.Artistic style, as the visual features of art works, represents the high level semantics of artifacts. Classification of artistic style depends mainly on human knowledge and experience. Due to the huge number of art works, classifying artistic style of art works by hand becomes infeasible. In this thesis, we proposed two artistic style classification schemes based on the SIFT features:the topic model and the pyramid match model. The pyramid match model:1) extract SIFT features of the painting, and let these low level features represent the style; 2) learn a codebook of the visual style, and use sparse coding to finding succinct representations for artistic style; 3) use pyramid matching as a way of measurement of the style similarity between art works; 4) train a linear Multi - class SVM model to predict which artistic style the art work belongs to. Different from the discriminative model, in order to complete the classification task, the topic model use the images’topic distribution getting from the topic learning procedure as the input to train the classifier.更多还原