【作者】 冯曦；

【导师】 康耀红；

【作者基本信息】 海南大学 ， 通信与信息系统， 2010， 硕士

【摘要】 为了更好地解决基于内容的图像检索中低级图像视觉特征和高级人类语言之间的语义鸿沟问题,在检索基本模块中加入相关反馈的模块,让用户参与到检索过程中,提交对于前次检索的意见帮助机器学习,更进一步了解用户的检索需求,更可能准确地“猜测”用户意图。本文在构建相关反馈的检索模型时,对图像进行分块提取特征向量,使用直方图提取颜色特征,使用基于MPEG-7的边缘直方图描述符提取纹理特征,使用多元正态分布来建模特征空间,使用LOGISTIC回归模型对图像特征向量的内部向量进行动态权值的调整,使用新的权值计算图像间相似度,使用贝叶斯估计模型估算出调整权值后图像数据库中图像符合用户前一次标记特点的概率,以此为排序基础对图像进行降序输出。主要研究工作有：1.学习了现有的图像特征提取方法,特别是颜色和纹理特征的提取算法,根据人眼对图像颜色和纹理的视觉感知敏感度的不同,提出分配两种特征不同维度的提取方法,更高效率地利用特征向量的维度。2.通过研究动态调整特征权值的相关内容,选择LOGISTIC回归模型计算特征向量内部向量的动态权值。3.进行过权值调整后采用贝叶斯估计模型,计算预测概率,用预测概率取代相似度作为输出图像的顺序。4.在matlab平台下提取特征,在Java环境下搭建仿真系统,经大量实验证明此方法有较好的查全率和查准率。平均查全率为0.80,平均查准率为0.85。更多还原

【Abstract】 To solve the semantic gap problem between low-level visual features in images and high-level human language in content-based image retrieval system better, it makes users easier to participate into the retrieval process by adding relevance feedback module to the basic retrieval module. When uses present judgment for previous retrieval result to help machine learning and better understanding the users’ intention, it will guess the intention more accurately.In this relevance feedback model, the feature vectors are extracted in five sub-blocks, the color features are extracted by histogram, the texture features are extracted by using edge histogram descriptor based on MPEG-7, the feature space is builded with multivariate normal distribution, the dynamic weights of the internal vectors are reckoned using LOGISTIC regression model, and finally Bayesian posterior is used to estimate posterior probabilities of all images which shows how the retrieved images after adjusting weights meet the users’ requirement. Correlation images are outputted according to the probabilities in descending order. The main research works are:1. Learning the current image feature extraction methods, especially color and texture feature extraction algorithm, extracting two different dimensions of two kind of features is more efficient.2. The LOGISTIC regression model used to calculate the dynamic weight of the initial feature vectors is choosen after searching for relevant content about dynamic adjustment of feature weights.3. After adjusting the weights, using Bayesian estimation model to calculate the predicted probabilities, images are exported on the basis of predicted probabilities instead of the sequence similarities.4. The experiments are done by matlab and Java softwares and the results show that this method has better recall and precision. The average recall is 0.80 and the average precision is 0.85.更多还原