【作者】 戎怡；

【导师】 胡正平；

【作者基本信息】 燕山大学 ， 信号与信息处理， 2010， 硕士

【摘要】 场景图像分类研究是对包含若干语义信息的图像集合进行分类的过程,可以对海量图像进行有效浏览与检索,成为当今计算机视觉研究领域的一个核心问题。鉴于图像与文本的关联性,将文本的词包模型与潜在语义分析模型运用到场景图像描述与分类上,具有重要的研究意义。针对当前图像特征提取算法有效性与复杂性相互制约的问题,展开以下研究:首先,构建了基于灰度图像局部边缘稠密采样区域的边缘改进局部二值模式(Edge Improved Local Binary Pattern,EILBP)特征,算法简单,性能稳定,能够对边缘信息丰富的图像进行合理描述,结合概率潜在语义分析(Probabilistic Latent Semantic Analysis,PLSA)模型完成场景分类,实验结果表明该特征提取算法应用在场景图像分类是有效的。然后,在EILBP特征的基础上,根据对称性构建了图像局部区域的边缘改进中心对称二值模式(Edge Improved Center Symmetric Local Binary Pattern,EICS-LBP)特征;针对彩色图像的颜色信息,构建了统计边缘主色对特征描述局部区域的边缘主色对信息;然后结合扩展PLSA模型完成场景分类,实验结果表明该方法具有较好的分类性能,对具有边缘轮廓的彩色图像分类精度高。最后,针对传统的视觉单词没有考虑特征间的依赖关系,不能充分表达图像主题这一问题,在彩色图像的EICS-LBP与统计边缘主色对特征的基础上,构造了一种含有上下文信息的视觉特征,之后结合扩展的PLSA模型实现场景分类。实验结果表明该方法具有较好的分类性能,对上下文信息丰富,具有边缘轮廓的彩色图像分类性能较好。更多还原

【Abstract】 The research of scene image classification is that how to make computer vision systems to classify the image sets which contain semantic information, according to understanding and discriminating the scene image of human. Scene classification is the core issue in the computer vision and image understanding research area, which could organize and process large mount of image data, and then used to retrieval or scan images reasonably or effectively. In consideration of the relationship between image and text, it is significant to make the bag-of-words model which is used to text corpus research area to describe the image and use the probabilistic latent semantic analysis model to classify the image. For the mutual restriction between the effectiveness and complexity of the image feature extraction algorithm, we employ the following research:First of all, we built the Edge Improved Local Binary Pattern (EILBP) feature of the local region which center point is formed by dense sampling the edge of gray image, it is simple, stable, and it can give a reasonable description about the gray image which contain rich contour information, and then we can obtain potential semantic of image by Probabilistic Latent Semantic Analysis (PLSA) model, after that we can accomplish the scene classification by K-nearest Neighbours Classier (KNN) classifier. The experiment results show that this method could achieve a higher accuracy, especially perform well in multi-edge gray images.Then, we construct the Edge Improved Center Symmetric Local Binary Pattern (EICS-LBP) feature of the local region which center point is formed by dense sampling the edge of gray image, it is produced based on the EILBP feature and the symmetry. For the color information of the color image, we construct the feature of statistical edge domain color pairs, it can describe the edge domain color pairs information of the local region. After that the new visual vocabulary is created by linear combination of the two species of visual vocabulary which are formed by clustering the corresponding features from dense sampling regions respectively. At last, we can obtain potential semantic by extended PLSA model and accomplish the scene classification by KNN classifier. The experiment results show that this method could achieve a higher accuracy, especially perform well in multi-edge color images.Finally, the traditional visual words method considers nothing about the reliance among of features, it couldn’t express the theme of image well. To overcome its defect, we propose the contextual visual features based on the EICS-LBP feature and the statistical edge domain color pairs feature of the color image. At last, we can obtain potential semantic by extended PLSA model and accomplish the scene classification by KNN classifier. The experiment results show that this method could achieve a higher accuracy, especially perform well in color images which contain multi-edge and rich contextual information.更多还原