Video Retrieval Based On Fuzzy Clustering And Image Features

With the increasing of video in TV program and Internet, it is more important to find the information in need quickly in digital library and VOD and long-distance education so that the technology of the content-based video retrieval turned into hotspot. Many people focus on the video structure analysis, such as shots division or key frame extraction. A method of content-based video retrieval plays a critical role in video retrieval, but it is very difficult and complex. This thesis discusses the method of video retrieval based on fuzzy clustering and image features. Methods and techniques to automatically analyze video documents based on contents and build structures that facilitate both hierarchical organization and search in video databases.At first, we summarize the key techniques used in the content-based video retrieval, such as shots division, video character analysis, shots clustering, etc.In fact, all video retrieval methods are almost based on image features. The retrieval process is divided into three phases by us: clips, episodes, frames. The comparability of clips depends on episode representation frames. We supposed subshots division as the first step, because shots division is complex. But subshots division needn't strict standard and easy to realize.We applied Shi Zhi-ping's method of multi-feature data association histogram such as color and texture in describing the image features of video frames and subshots.We proposed a new video retrieval algorithm based on fuzzy clustering and image features. Firstly, we brought forward a new approach for selecting episode representation frames based on subshot clustering by using C-Mean fuzzy clustering algorithm. Secondly, we proposed the method of video retrieval based on representational frame sequence retrieval. The experiment shows that we could get better result in video retrieval.We bring out our video retrieval method based on multi-feature data association histogram and C-Mean fuzzy clustering algorithm. No complicated structures are involved in our method. At the same time, it is more accurate with multi-feature comparison. It is also robust to holes of boundaries measurement.