【作者】 田刚；

【导师】 王天江；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2006， 硕士

【摘要】 随着信息技术的高速发展,数字图像急剧膨胀。如何快速提取感兴趣的目标图像,成为海量信息处理面临的瓶颈。聚类分析近十年来发展迅速,被广泛应用到人工智能、信息控制、医疗诊断、天气预报、图像分析等领域,其中使用最普遍的是C-均值聚类。C-均值聚类虽然在图像处理中应用比较广泛,但也存在很多不足。基于此,在聚类基础上,引进相关反馈机制对聚类进行调整,称之为学习聚类。当前主要的图像检索方法有基于内容的和基于语义的。这两种方法各自存在优缺点,同时也存在一些共同的难点。结合两个图像检索算法的特点,在学习聚类的基础上提出了基于底层特征和语义的图像检索算法BFSR(Based on Feature and Semantic Retrieval Algorithm)。它存在不少优点:有效的解决了初始类中心的选取、能自动地进行语义标识、能从多个角度表达图像内容,半自动地确定聚类数目等。但也存在检索精度不高,无法解决语义的同义性和分歧性等问题。存在这些问题的主要原因是由于侧重点不同,用户对图像相似性判断也存在不同的标准。基于此,在BFSR的基础上结合相关反馈策略改进得到了基于相关反馈的学习聚类算法RFCL(Based on Relevant Feedback Cluster Learning Algorithm)。RFCL能够根据用户行为猜测用户的兴趣,并能自动地调整相似性度量准则来提高检索的准确率。根据RFCL的思路开发了一个界面设计合理的试验系统。通过对查全率和查准率的分析,显示RFCL具有能消除语义的同义性和分歧性、有效提高检索精度等性能。更多还原

【Abstract】 With the development of information technology swiftly, data images become more and more popular. How to get the interest image from much image quickly is a key problem. Cluster analysis develops quickly, and is applied to artificial intelligence, information control, medicine diagnose, weather forecast, mage analysis and so on. The c-means cluster is most popular in all clusters.The c-means cluster is applied image analysis, but it has much shortcoming. Based on that, feedback is added into cluster, so that cluster can learn self. This cluster based on feedback is called learning cluster. Currently, three are two main image retrieval methods. One is based on content, another is based on semantic. Both have different merit and shortcoming, and have the same difficulties. BFSR(Based on Feature and Semantic Retrieval Algorithm) is brought forward, which combines traits both retrieval methods and is based on learning cluster. It has much merits, like figuring out initial center choose, label semantic automatically, depicting image content in many points of view. But it has a lot of shortcoming, like bad precision, semantic synonymy and bifurcation. Based on which, RFCL(Based on Relevant Feedback Cluster Learning Algorithm) is brought forward, it increases precision of result effectively.An experiment system is developed based on RFCL. The result reveals that RFCL can clear up synonymy and bifurcation of the semantic, and increases nicety of result via analysis recall and precision.更多还原