Research On Some Key Issues In Content Based Video Image Segmentation And Retrieval

With the rapid development of multimedia communication technology, large multimedia data has been gradually becoming the focus object in information processing and transporting, especially video image data. In many application area, millions bits of video data is generated everyday; in the particular application areas, such as video tracking, image retrieval and so on, it needs to do fast segmentation, feature extraction and recognition on video image; due to the large data, it is very difficult to represent, store, transport, and recognize the data, and it is impossible to retrieval these information if the data couldn’t be used in rational structures, and making this to be most challenging problem in current research. Content-based Image Retrieval technology is a key to solve this problem.In this paper, analysis the current features of video segmentation technology, that is separating image into several independent areas and extracting fields of interest, but it is very difficult to do segmentation under the lack of sufficient previous information, especially in motion sequences, the relation between two frames needs to be considered, due to the sensation of boundary, it has some problems in lack or over segmentation. Currently, there exist two popular methods:semi-supervised and unsupervised, in semi-supervised condition,, the class of image is defined by users, but in unsupervised condition, the number of classes is defined by system automatically, and improving difficult on the other hand. To video image retrieval technology, it is very important in real time processing in large database, and has high quality in speed and accuracy, the main technologies are key frame extraction, object features extraction, similarity measures, quarry function and so on. In this paper, several problems in CBIR system have been done depth research, and made some results as follows:Firstly, in the problem of feature extraction and classification from image and sequence, Markov Random Fields based image segmentation is proposed for the inefficiency of single feature and description. Combining texture and color features and using MRFS models to separate object; when comes to bad analyzing the tendency of moving object by differential algorithm between two frames, we combined motion estimation and proposed a affine motion model and MRFS based feature classification model, and realized the segmentation in image and video sequences.Secondly, in the aspect of efficient feature extraction and retrieval in image quarry, in this paper, we did depth research in characters of wavelet transform, and proposed M band wavelet based image feature extraction, using clustering technology in classification and similarity computing in image matching to output the result. Under the relevance feedback method, we proposed a multi-layers structure and combining spatial and frequency domain features for image retrieval.Thirdly, in the area of special object detection and recognition in video retrieval, we analyzed the application of statistical features in face recognition, and model face and nonface features and then combined affine motion model in sequences, and proposed a Bayesian classification and Support Vectors Machines classification optimal algorithm in face detection and recognition in videos, and the experiments show the efficient of this method.Finally, in the design of data structure in video retrieval system, we analyzed the characters of video information, used mutual information to extract key frames in sequences; we proposed a metadata information features and storing technology, and gave details in generating metadata and structuring processing, and finally we proposed a method in face image video retrieval based on combination of face features and metadata technology.