Study on the adaptive high-dimensional information retrieval based on learning and its applications

基于学习的自适应高维信息检索及其应用研究

• 批准号: 18500110
• 负责人: MARUYAMA Minoru
• 金额: $ 1.83万
• 依托单位: Shinshu University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 2006
• 资助国家: 日本
• 起止时间: 2006 至 2007
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-18500110/
• 关键词: learning; probabilistic tonic models; EM algorithm; image classification; image segmentation; image retrieval; image annotation; pLSA; LDA; パターン識別; SIFT; クラスタリング; 確率モデル; SVM; 混合正規分布モデル

In this research project, we investigate the content- based image analysis via learning probabilistic generative model from examples Nowadays, there exists various kinds of media information, such images, movies, on the network 'lb efficiently access to such media information, the method of content-based information retrieval is required. 'lb realize the method, in this research, we deal with the content-based image categorization, segmentation, retrieval and image annotation. For that purpose we exploit the probabilistic topic models. In the models latent topic variables am introduced to represent the content of the image. Among the probabilistic topic models, we mainly use mixture model, pLSA (probabilistic latent semantic analysis), LDA (latent Ditichlet allocation) for image analysis. For learning model parameters from examples, we use EM algorithm (for mixture model and pLSA) and variational methods (for IDA). Based on the topic models, we propose methods for image categorization based on the estimation of the conditional probability p(category I image), bottom-up image segmentation for document images, and content-based image retrieval and image annotation based on the quay words. For image annotation and image retrieval by query word, probabilistic model of image and keywords is required. We propose several models for the task. We examine the effectiveness of the proposed methods by experiments using image database, including Caltech101 database and MIT LabelMe database.

在本研究项目中，我们通过学习概率生成模型对基于内容的图像分析进行了研究。目前，网络上存在各种媒体信息，如图像、电影等，有效地获取这些媒体信息，就需要采用基于内容的信息检索方法。本文实现了基于内容的图像分类、分割、检索和图像标注。为此，我们利用了概率主题模型。在模型中，引入了潜在主题变量来表示图像的内容。在概率主题模型中，我们主要使用混合模型、概率潜在语义分析(PLSA)、潜在Ditichlet分配(LDA)进行图像分析。为了从实例中学习模型参数，我们使用EM算法(对于混合模型和pLSA)和变分方法(对于IDA)。在主题模型的基础上，提出了基于条件概率p估计的图像分类方法、基于文档图像自底向上的图像分割方法、基于码头词的基于内容的图像检索和图像标注方法。对于图像标注和基于查询词的图像检索，需要图像和关键词的概率模型。我们为这项任务提出了几个模型。通过对Caltech101图像数据库和MIT LabelMe图像数据库的实验，验证了所提方法的有效性。

项目成果

An online handwritten music symbol recognition system

在线手写音乐符号识别系统

• 发表时间: 2007
• 期刊: International Journal on Document Analysis and Recognition 9
• 作者: 山田悦明;原田利宣;精密工学会・デジタルスタイルデザイン研究分科会;デジタルスタイルデザイン研究分科会;H. Miyao and M. Maruyama
• 通讯作者: H. Miyao and M. Maruyama

An online music symbol recognition system

在线音乐符号识别系统

• 发表时间: 2007
• 期刊: International Journal on Document Analysis and Recognition 9・1
• 作者: Hidetoshi Miyao;Minoru Maruyama
• 通讯作者: Minoru Maruyama

Feature extraction for document image segmentation by pLSA model

pLSA模型用于文档图像分割的特征提取

• 发表时间: 2008
• 作者: Takuma Yamaguchi;Minoru Maruyama
• 通讯作者: Minoru Maruyama