Researches On Manifold Learning Based Image Retrieval And Its Application

With the rapid developing of multimedia and internet technology, people now have all kinds of ways to communicate with other person at any time and wherever he is. They can chat, glance at videos, browse a variety of webpage and look for pictures what we need on line and so on. And these methods which we use to interchange are constantly updated and renovated by people along with the developing technology. Through the above said, our lives are more and more comfortable and convenient. Such as we can know everything occurred in the world without leaving the house, and if we want to view videos and photographs in the Internet, only with gently clicking the mouse, we can find thousands of videos and photographs. But growing numbers of videos and photographs bring some negative aspects, for example the waste of information sources. Our lives are filled with large scales of information; on the other hand, we are lack of intellectual. When we extract useful and reasonable data from the huge amount of data sets, how to maintain the completeness of data information which satisfies person’s storage requirement and need-apperceived is a question that we should solute immediately. As an unsupervised and statistical learning method, manifold learning gives us a train of thought.When we process various kinds of data, samples that we need increase exponentially; the distances among these samples are smaller and smaller, this is the curse of dimensionality what we say now. Fortunately, a lot of questions in practice, for example, most of observation data variables in high-dimensionality can be expressed by a few of influencing factors. This states that most of observation data variables in high-dimensionality conclude large amounts of data’s redundancy, and usually there is strong relevance among each component of the data. This phenomenon represents as data locating in low-dimensional manifold or near the low-dimensional manifold in geometry. In order to effectively reveal data’s potential structure, we should learn and find out its low-dimensional feature which is embedded in high-dimensional space. This is the main purpose and theme of manifold learning.During recent years, manifold learning method was used in fields of pattern recognition, computer vision etc at first. And now it is expanded to content based image retrieval (CBIR). face recognition, video copy detection, and medical image processing, and shot segmentation and so on. This paper mainly introduces the application of manifold learning in CBIR and shot segmentation. Because a shot is made up by frames which have similar features, CBIR method based-on manifold learning is tried to apply in shot segmentation. Features extracted from video frames and images are dimensional-reduced to low-dimensionality. And data in low-dimensional are calculated the distances by Euclidean. Finally, the distances are compared with the threshold. If the distances are in the control of the threshold, it states that it belongs to one shot or is similar to query image, or it belongs to different shot and is not similar to query image.The main innovation and contribution of this paper are:(1) It introduces some manifold learning method, and connects them with content based image retrieval. One of the essential questions CBIR techniques is:image is represented by feature vectors, ordinary they are high-dimensional vectors. So in large data base, when it is compared by sequence, the process is time-consuming. The main idea of manifold learning method is to discover the essential rule among high dimensional data and linearly or nonlinearly map the data from high dimensional space into low dimensional one to get the compact low dimensional representation of the original high dimensional data. Connecting the manifold learning and CBIR can efficiently solute the problem that when image feature vectors queried in image data base is time-consuming and heavy computation.(2) It introduces shot segmentation method based on manifold learning. As the video blog is becoming popular, video processing and analysis receive increasing attentions. Video shot segmentation is the fundamental processing of video analysis, such as key frame extraction, video summary, video retrieval, and so on. So extracting efficient and reduced feature is necessary. The idea of combining manifold learning and video shot segmentation can accomplish is feasible in practice. As the high-dimensional vectors, video frames are dimensionality reduced by manifold learning method.