Video retrieval and classification is important in multimedia application in the future. Motion information, as the unique feature of video, is essential in the research of video retrieval. Generally speaking, video motions can be divided into two types, global motions and local motions. The former is caused by camera movements, and the latter refers to motions of objects in the scene. Traditional retrieval schemes are designed for just one type of video. They obtain good experimental results in certain video application domains. However, if video with complicated motions are applied, or if motion types of shots are not the type that the systems deal with, or there are coding errors, they are unreliable if. In the paper, motion information is analyzed with data mining theory, and a robust video classification method is proposed to make video retrieval more effective.In the paper, with the introduction of data mining theories and video codec standards, we propose a roughset based video pre-classification method, and design a motion-information-based video retrieval prototype system. Video pre-classification is to extract feature attributes of motion information and to obtain classification rules which are used to classify frames of video. The result of video pre-classification is further used for shot classification. The video retrieval prototype system is based on the method. It retrieves shots in the database with the same motion type as the submitted shot. It's able to avoid retrieval errors caused by motion prediction errors and unexacting processing results in the high layer. So the motion-information-based video retrieval can be more effective. The simulations demonstrate the roughset based video pre-classification is perfect. The retrieval results in the video retrieval prototype system also shows that the system can exactly retrieve shots with similar motion types. As a result, there is certain practicality of the system. |