The Research On Key Frame Selection And Feature Matching In Video Retrieval

As the Internet technology develops rapidly, various kinds of video websites have emerged, efficient organization and usage of video information turn out to be something that matters. Compared with the traditional text-based information, the sizes of the videos increase greatly and the structures become more complicated. As a matter of this fact, content-based video retrieval technology was quickly proposed and becomes the hot topic of the related researches. Content-based video retrieval refer to a special process in which users search what they need from a large amount of videos on the Internet or their PCs, with just a short query video.This paper discusses the critical technologies of related contents and key points, including shot boundary detection, key frame selection and feature extraction and feature matching. On these topics we all make further improvement. The main contents are as follows:Firstly, on shot detection, common shot transition types and the corresponding detection methods are discussed. A shot detection method based on frame difference proposed using color histogram and local binary pattern descriptor as image features to calculate the inter-frame difference. This kind of algorithm is simple and fast which not only the global color information is considered but also the local texture information.Secondly, on key frame selection methods, the principle of key frame selection is firstly introduced and then some key frame selection methods are discussed. A modified key frame selection algorithm is proposed based on the previous discussion. It turns out to be an effective key frame selection method following the shot boundary detection. The selected key frames turn out to be very representative among all the frames in a single shot. The redundancy of the details is greatly decreased and the key frame numbers are flexible.Finally, through hard working and general analysis on various kinds of image features, an algorithm of image feature extraction based on color histogram and texture feature is proposed. A good algorithm of image matching according to the special structure of the video is also proposed. The result of the experiment works out well to show that it has a degree of high matching accuracy and runs fast.