Basic Study on Image Retrieval in Super-Distributed Environment

本课题的研究内容包括超分布环境下图像信息（目标图像）的搜索方法和基于图像特征的图像检索方法，首先研究了目标图像在数据库中的位置概率先验已知的情况下图像信息的搜索方法。假设给定了搜索每个数据库所需的通信成本和时间。提出了使估计的通信费用和时间最小的搜索方法。数学分析表明，即使目标位置的先验概率存在一定误差，该方法仍能获得比随机搜索更好的性能。其次，研究了两种基于图像特征信息的图像检索方法，即基于特征信息的图像检索方法和基于特征信息的图像检索方法 关于我们 omatically.一种是“交互式方法”，另一种是“关键字方法”。在交互式方法中，系统向用户展示多个图像示例，并且用户选择与目标图像最相似的图像。系统根据用户的选择显示其他图像。经过反复的显示和选择，系统知道目标图像中包含的特征信息，并进行检索，实现了交互式方法的原型，目前该原型已经可以通过Internet访问。在关键字方法中，我们研究了从关键字到特征信息的自动转换。通过学习过程获得每个关键词对应的特征匹配规则和参数，并存储在字典中。采用遗传算法求解参数。初步研究表明，在允许一定误差的前提下，关键字转换为特征信息是可能的。少

This research project includes the searching method for image information (target image) in super distributed environment and image retrieval based on image feature.First, we have studied the searching method for image information in the case that probability of target image location in each database is a priori known. The communication cost and time which is required to search each database is assumed to be given. We have proposed the searching method which makes the estimated communication cost and time minimum. According to the mathematical analysis, it is shown that the proposed method gives better performance compared with random search even if a priori probabilities of target location contains some error. Further, we have developed the prototype system of agent which has capabilities to correct the information concerning the images stored in the network.Second, we have studied two types of image retrieval methods based on feature information which can be extracted from images aut … More omatically. The one is called "interactive method" and the other is called "keyword method". In the interactive method, the system shows the user plural number of image examples, and the user select the image which is most similar with the target one. The system shows the other images based on the user's selection. After the repetition of display and selection, the system knows the feature information which is included in the target image and retrieves it. We have made the prototype of the interactive method and now the prototype is available through Internet. In the keyword method, we have studied the automatic conversion from keyword to feature information. The rules and parameters which is used in feature matching corresponding to each keyword is obtained through learning procedure and is stored in the dictionary. Genetic algorithm is used to obtain the parameter. The basic study shows that it is possible to convert keyword to feature information if a little conversion error is allowed. Less