Extraction Of Video Fingerprint Based On Features Of Images And Audios

The video information is an extremely important part of the internet multimedia for which is exploding,while the hardware of internet has stepped into a slow development stage.It has brought series of problems with the coming of the huge amounts of videos,such as the unhealthy videos,pirated videos,illegal videos and the problems of retrieving on huge amounts of videos.In order to solve these problems,the video fingerprint technology is an efficient method than operating by manual detection and retrieval.The video fingerprint technology started in 1990s.The video fingerprint is the unique information represented for the video content that extracted from the video features.The main existing video fingerprint technologies are analyzed in this paper.Two video fingerprint algorithms based on the image-audio feature fusion are proposed to solve the problem that the mainstream algorithms can not get both the robustness,accuracy and real-time performance at the same time.The main works are listed as follows.First,the SIFT feature with good performance of robustness and accuracy was chosen as the image feature of the video fingerprint.The SIFT algorithm was improved to solve its defect in accuracy and real-time computing in the following parts.The I frames in MPEG video streams were selected as the key frames of videos,which can reduce the computational time and keep essential information of video at the same time.Then a pretreatment method of image cut was proposed in order to improve the defense ability against the icon and subtitle attacks.At last,the feature points were clustered to select the typical points for purifing the feature points.The iterations of RANSAC were reduced to improve the accuracy of video matching,which can increase the algorithm speed as well.The experiment results showed that a 15%higher precision rate of video matching was obtained through the proposed algorithm than the original SIFT algorithm,and the video matching computational time was 10%less than the original SIFT algorithm in all test time length.The proposed algorithm was robust to the brightness and size attactks.Second,the audio fingerprints extracted from the series sub-binds were chosen as the audio features of video fingerprints in this paper.Then the process of generating each audio fingerprint with 32 bit length was expounded.It consists of four parts including the audio encoding,audio frequency domain converting,frequency domain filterling and feature extracting.The chosen audio fingerprint was proved to be performing well in the real-time computing and accuracy.Third,two image-audio feature fusion algorithms were proposed for solving defects in the existing mainstream image-audio feature fusion algorithms.The image-audio fusion algorithm based on dynamic weight distribution adjusted the weighted value by the results of single feature fingerprint matching.The image-audio fusion algorithm based on Adaboost trains some videos to get a scientific weighted value with the Adaboost classifier.The experiment result showed that higher precision rate of video matching was obtained though both two proposed algorithms than the algorithms based on one single feature.The image-audio fusion algorithm based on Adaboost performed better than the dynamic weight distribution algorithm,and a 38%higher precision rate at 95%recall rate was obtained through The image-audio fusion algorithm based on Adaboost than the SIFT algorithm.