19-10-2012, 03:31 PM
A Hypothesis Test based Robust Technique for Video Sequence Matching
A Hypothesis Test based Robust Technique.pdf (Size: 1,023.28 KB / Downloads: 29)
Abstract
Video sequence matching is the most crucial step to verify whether a video sequence has been
copied from another or not. The video sequence in question is represented by a set of visual
descriptors. Based on such descriptors the sequence has to be matched with the sequences
present in the database. Thus, sequence matching becomes a crucial task. Furthermore, to
evade the verification process, one may incorporate visual distortions in the original
sequence to generate the copied version. Again, the copied sequence may be generated by
taking few sampled frames or a part of the original sequence. Thus, the task of sequence
matching becomes more challenging. In this work, rather than concentrating on the visual
descriptors, we have focused on the technique of sequence matching that can sustain the
challenges posed. We have proposed a hypothesis test based scheme for sequence matching
which has inherent strength to withstand the said attacks considerably. Experimental result
also indicates its capability of the scheme in doing so.
Introduction
Technological development has made capturing and storage of video data easier and
inexpensive. Moreover, development in the arena of network and communication and
increase in bandwidth has encouraged video sharing, broadcasting enormously. Availability
of digital videos on various media like TV-channels, web-TV, video blogs, public video web
servers has led to huge growth in video data volume. All these have an adverse effect on
copyright management. The technology has enabled easy access, editing and duplicating of
video data. Such activities result into violation of digital rights. Considering the huge volume
of the video database, controlling the copyright of videos generated everyday has become a
critical issue. But, it is the basic requirement in protecting the intellectual property right.
Driven by the importance of copyright protection, a new area of research called video
fingerprinting has come up. Lee et al. [1] has defined fingerprint as perceptual features for
short summaries of a multimedia object. The goal of video fingerprinting is to judge whether
two video have the same contents even under quality-preserving distortions like resizing,
frame rate change, lossy compression [2].Video fingerprinting is also commonly referred to
as video copy detection.
Past Work
Features of a video copy detection system must satisfy the properties outlined in [2].
It must be robust so that fingerprint of a degraded video and the original one should be
similar. It should be pair wise independent to have different fingerprints for
perceptually different fingerprints. Finally, the fingerprint must support fast search i.e.
it should be search efficient.
Proposed Methodology
In a video copy detection method, the task is to verify whether or not a test/query
sequence is a copied version of a sequence present in the database. It has already been
discussed that such a system consists of two major modules namely extraction of
signature (feature vector) and sequence matching. Signatures must fulfill the diverging
criteria such as discriminating capability and robustness against various geometric and
signal distortions. Sequence matching module bears the responsibility of devising the
match strategy and verifying the test sequence with likely originals in the database. It is
evident from the past work that to achieve robustness of the detection system, the
emphasis has been put mostly on the development of attack invariant features. For
matching the sequences, the researchers have mostly relied on certain threshold based
comparison of feature vectors. Thus, a robust matching technique capable of sustaining
the attacks can further enhance the capacity of a copy detection system. In this work,
we put our effort in developing the sequence matching module which will have its own
inherent strength to combat the attacks adhered by copier. We have relied on hypothesis
test based strategy as presented in [35] for the purpose. As the video signature is likely
to be multi-dimensional, we have considered multivariate Wald-Wolfowitz run test [36]
based hypothesis testing.