Video Phylogeny based on Fingerprint Features for Near- Duplicate Video Clips Detection and Parent tree generation

Abstract

Video Phylogeny is the task of reconstructing the ancestral tree among the pair of near duplicate

videos. Due to the huge availability of various tools and software's, significant amount of

available video content was reused by many social web sites such as face-book and you-tube.

Various transformations are applied on these video content e.g. rotation, scaling, contrast and

background modification. We have introduced fast and durable fingerprint extraction and

matching algorithms based upon Tree-lets, Wavelet Energy and Gabor over Wavelet for various

types of video transformations and then reconstructed a video phylogeny tree based upon

threshold values of the resulting hash generated. The phylogeny tree further classifies these

videos into various types based upon transformations applied say affine transformed, scaling,

background and sub-scene changes. The fundamental goal of our method is to extract signatures

or feature vector (color, edge based, temporal etc.) from video to determine and detect whether it

is an authorized version or unauthorized version of the parent video. The proposed technique

depends upon spatial and edge based features of video frames and is validated with real world

data sets downloaded from you-tube. Moreover, the resulting fingerprint requires reduced

number of bits and do not require extensive storage space being small in size. Our method has

demonstrated precision, accuracy and robustness against various types of video transformation

techniques.