| In the last decade, along with the rapid development of informationtechnology, more and more multimedia information is available, especiallyfor video data. The proliferation of digital videos has made accessibility ofvideo contents easier and cheaper while being the source of many problems,e.g., how to get the interested video from the vast amount of video data. Atthe same time, it becomes more complicated for content management becauseof the huge and ever increasing amount of videos. Keywords-based traditionaldistinguishing and retrieval method cannot work efficiently anymore, it isimportant to find a new method for video content discrimination, retrieval,monitoring services and authentication. Video fingerprinting is a techniquethat can be used to solve these problems and for this reason it has become ahot research topic in recent years.A video fingerprint is an identifier that is extracted from a piece of videocontent and discriminates one video from the others. The process of extractinga fingerprint from the video content is referred to as fingerprinting the videoor video fingerprinting. Pereptual video fingerprint means if two video clipsare perceptually different, the fingerprints extracted from them should beconsiderably different, and if two video clips are perceptually samilar, thefingerprints extracted from them should be considerably identical.This paper focuses on research about feature extraction, fingerprintmodeling and similarity matching of video fingerprinting. Three digital videofingerprint algorithms are presented in this paper.1. A fingerprint algorithm based on relative orientation invariant betweengeometric centroid is suggested. The perceptual centroid means that the sameperceptual content video should have the same centroid, further, theorientation of the centroid is identical, and the relative orientation keeps unchangeable. In this method, the geometric center of every frame is used asoriginal point and the orientation of centroid is calculated based on luminancegeometric centroid. Converting orientation from the original frame rate to afixed frame rate through temporal distance and a new orientation vector isgenerated. The next, an invariant orientation is created through successive twoorientation subtracting. At last, video fingerprint is generated byconcatenating all of the invariant orientations. The experimental results showthat the proposed algorithm is robust against lossy compression, frame ratechange, resizing, rotation, cropping, global change in brightness and gaussianwhite noise. The quality evalution using structural luminance comparisonimproves the visual effect.2. Presented is a new video fingerprint base on mean, instead of usingtransform for further processing. Usually the video fingerprint only candiscriminate one clip from others, but it is difficult to assess the video qualitybetween the identical content video. The input video is firstly converted into anew video with a fixed frame and a fixed size. Secondly, the new video ispartitioned by a macroblock(16×16). At last, fingerpring is formed bynormalizing the luminance mean that is calculated from every macroblock.The experimental results show that the proposed algorithm is robust againstlossy compression, resizing, frame rate change, global change in brightnessand gaussian white noise.3. Wavelet transform is usually used in digital image processing, becauseit can extractes some characteristic parameters in this frequence domain. Acomplex wavelet is constructed based on a low-pass filter and the discretewavelet coefficients are achieved through sampling of the continuous wavelettransform in this video sequence. The discrete wavelet coefficients are used asvideo fingerprints. The proposed algorithm is robust against resizing, smallrotation. The experimental results show that the proposed algorithm has a lowfalse negative rate and false positive rate.
|
PDF Full Text Request