INTERACTIVE RETRIEVAL AND FILTERING TECHNIQUES FOR VIDEO DATA

本研究的目的是视觉媒体的相似性搜索。我们开发了一些个性化检索系统的基本技术，如基于相关反馈的交互式检索方法、基于过滤检索方法的加速技术和基于协同过滤技术的用户偏好处理方法。首先，我们开发了基于边缘方向直方图的相似度搜索、基于图像特征点提取的图形直方图检索和基于特征向量之间的双线性形式相似度的图像检索技术。接下来，我们将这些图像检索方法应用于视频帧的相似性搜索。在发展了帧数混合分布视频的表示方法之后，我们首先提出了一种基于混合分布距离的视频检索方法。接下来，我们提出了一个视频的特征向量，称为时间相关四边形，它表示连续视频帧之间的相关性，并利用该时间相关图开发了视频的相似性搜索技术。我们还开发了一种基于潜在语义分析的视频相似度搜索方法，通过视频镜头配置的列联表来表示视频数据库。为了提高该视频检索方法的性能，我们开发了一种基于潜在语义分析的鲁棒聚类方法。为了加速基于这些方法的图像或视频的搜索，我们开发了一些滤波搜索技术，如基于特征向量间双线性形式距离下界的滤波、基于时间相关图上界的滤波、基于图直方图下界的滤波和基于直方图交点上下界的滤波搜索技术。我们研究了根据用户对视频的预测偏好为用户推荐视频的协同过滤方法。我们开发了一种基于聚类用户的加速协同过滤方法，一种基于项目之间相关性的偏好估计快速协同过滤方法，以及一种选择最优项目集向新用户展示的方法。少

The aims of this research is the similarity search of visual media. We have developed some basic techniques for personalizing retrieval systems such as interactive retrieval methods based on relevance feedbacks, speeding-up techniques by filtering retrieval methods and user preference handling methods based on collaborative filtering techniques.At first, we have developed some image retrieval methods such as the similarity search by using edge orientation histograms, retrieval with a graph histogram based on extraction of image feature points and an image retrieval technique using bilinear form similarity between the feature vectors.We have next applied these image retrieval methods to the similarity search of video frames. We developed firstly a video retrieval method based on the distance between mixture distributions after the development of the representation method for videos with mixture distributions of frames. We next proposed a feature vector of videos called the temporal corr … More elogram which represents the correlation between consecutive video frames and by using this temporal correlogram we have developed a similarity search technique for videos. We have also developed a similarity search method of videos based on the latent semantic analysis with the representation of the video database by a contingency table of shot configuration of videos. We have developed a robust clustering method for the improvement in the performance of this video retrieval method based on the latent semantic analysis.At the aim of acceleration of the search of images or videos based on these methods, we have developed some filtering search techniques such as the filtering based on the lower bound of the bilinear form distance between feature vectors, the filtering based on the upper bound of the temporal correlogram, the filtering based on the lower bound of the graph histograms and the filtering search technique based on the upper and the lower bounds of the histogram intersection.We have examined the collaborative filtering method for recommendation of videos for users according to their predicted preference of videos. We have developed an acceleration method for the collaborative filtering by clustering users, a fast collaborative filtering method of preference estimation based on the correlation between items, and a method for selecting an optimal set of items for presenting them to a new user. Less