Watermarking of MPEG-4Videos
Abhinav Gupta and Phalguni Gupta
Department of Computer Science and Engineering
Indian Institute of Technology,Kanpur,India-208016
{abhigupt,pg}@cse.iitk.ac.in
Abstract.A MPEG-4compressed domain video watermarking method
is proposed and its performance is studied at video bit rates ranging
from64Kb/s to900Kb/s.The watermark is inserted by modifying
Discrete Cosine Transformation(DCT)coefficients.The strength of wa-
termark is changed according to local frame characteristics to reduce
impact on visual quality.The algorithm’s performance is also studied for
watermarking bits in a frame ranging from1Kb/frame to3Kb/frame.
The watermark is attack-free against attacks like scaling,rotation and
cropping even if blind-techniques are used.
1Introduction
Digital Watermarking is imperceptible insertion of information into multimedia data which act as a signature in the video.The classical approach to watermark-ing of a compressed video stream is to decompress the video and use a spatial-domain or transform-domain watermarking technique and then recompress the watermarked video.Some of the disadvantages of this approach are:noise ad-dition due to recompression,bad quality due to no knowledge of compression parameters results and computational complexity.Another approach is to insert the watermark in compressed domain only.In this approach the watermark is inserted in some syntactical elements like DCT coefficients of a partially decoded video.
Hartung[1,2]describes techniques to embed a spread-spectrum watermark into MPEG-2[3]compressed videos as well as into uncompressed video.For compressed domain watermarking t
hey decode the video to obtain the DCT coefficients of each frame and insert the watermark by modifying those DCT co-efficients.Langelaar[4]describes a compressed domain watermarking technique called Differential Energy Watermark(DEW)in which the video is partitioned into groups of blocks each of which is further divided into two sets of equal size as determined by the watermark embedding key.By comparing the energy of selected DCT coefficients within two sets,a single payload bit is expressed.Jor-dan et al.[5]proposed a method for the watermarking of compressed video that embeds information in the motion vectors.Hsu and Wu present a method for watermarking[6]which modifies middle frequency DCT coefficients in relation to spatially and temporally neighbouring blocks.Nicholson[7]evaluated the wa-termark robustness and video quality after video is watermarked and compressed D.Zhang and A.K.Jain(Eds.):ICBA2004,LNCS3072,pp.746–752,2004.
c Springer-Verlag Berlin Heidelberg2004
Watermarking of MPEG-4Videos747 with MPEG-4.However none of these techniques addressed direct watermarking of these MPEG-4videos.
In this paper,a new compressed domain watermarking technique for MPEG-4[8]video streams is pre
sented.The approach is similar to[1]as watermark is inserted in DCT coefficients.However use of synchronization templates makes it robust and local gain method improves the quality.
In Sect.2an overview of MPEG-4technique is presented.This is followed by the proposed method and results in Sect.3and Sect.4respectively.Sec-tion5proposes a protocol for subjective evaluation of the system and the paper concludes in Sect.6.
2Overview of MPEG-4
MPEG-4[8]encodes the visual information as objects(natural,synthetic video and still textures).MPEG-4encodes a description of the scene for proper ren-dering of all objects.An MPEG-4visual scene may consist of one or more video objects.Each video object is characterized by temporal and spatial information in the form of shape,motion and texture and corresponds to a2D object in the scene.A Video Object Plane(VOP)is a time sample of video object.VOP’s can be encoded independent of each other or dependent on each other by motion compensation.A VOP contains the encoded video data in form of macroblocks.
A macroblock contains a section of the luminance component and the spatially subsampled chrominance components.In MPEG-4visual standard there is sup-port for only one chrominance for
mat for a macroblock,the4:2:0format.In this format,each macroblock contains4luminance blocks,and2chrominance blocks.Each block contains8x8pixels encoded using DCT transformation.The DCT coefficients are then adaptively quantized to achieve low bit rates.
3Proposed Method
In the proposed method,a watermark signal is inserted directly into MPEG-4 compressed bit-stream while detection is performed using compressed bit-stream without any watermark signal.Section3.1discusses the formation of water-mark signal from the original message signal and synchronization templates. Section3.2addresses the process by which the watermark signal is embedded in the MPEG-4videos.Section3.3discusses the local adaptive gain method to increase quality of the videos.
3.1Spread Spectrum Watermark Signal
The watermark signal is often limited to a small value to ensure the imper-ceptibility and subject to interference from the host signal and additional noise arising from subsequent processing.A spread spectrum signal is vulnerable to synchronization errors occurring after scaling,cropping and rotation.A pair of
748Abhinav Gupta and Phalguni Gupta
templates is imposed on the spread spectrum signal to combat synchronization losses.Thefirst template restricts the watermark signal to have a regular pe-riodic structure.In particular,the watermark signal w(x,y)is constructed by repeating an elementary watermark tileˆw(x,y)in a non-overlapping fashion.If the tiling is done properly the peak always occur at the centre of each tile.If a linear transformation A is applied to watermarked VOP,the autocorrelation coefficients h(x,y),thus the new peaks move to x’and y’according to
[x y ]T=A[x y]T(1) The second synchronization template forces w(x,y)to contain a constellation of peaks in frequency domain.This requirement is met by makingˆw(x,y)as combination of the message bearing symbol m(x,y)and the synchronization signal g(x,y).In frequency domain this g(x,y)contains peaks in mid frequency band each peak occupying one frequency coefficient and having unity magnitude. After the geometrical transformation A applied to image the FFT coefficient F(u,v)move to a new location(u ,v )according to the equation
[u v ]T=(A T)−1[u v]T(2) Thus using the two equations above we can obtain the linear transformation A applied to the image and hence revert the transformation to perform the detection method.Figure1(a)outlines the process of creation of watermark signal.
3.2Watermark Embedding
ignore subsequent bad blocksThis section describes embedding the watermark directly to the bit-stream gen-erated in accordance with the Advanced Simple Profile(ASP)of the MPEG-4 standard.The watermark signal w(x,y)is added to the luminance plane of the VOPs.Since the DCT is a linear transform adding the transformed watermark signal directly to DCT coefficients of the luminance blocks is equivalent to ad-dition in spatial domain.An elementary bit-stream is parsed down to the block level and variable length coded motion vector and DCT coefficients are obtained. Motion vectors are reconstructed using VLC decoding and reversing the predic-tion steps wherever applicable.After the watermark signal is embedded,VLC codes are regenerated and bit-stream is reconstructed.Fig.1(b)depicts the whole outline.
Since a96x96signal is embedded in a192x192image,a8x8block is embedded
in16x16block.Hence,the total number of permutations possible are P256
64.This
property helps to introduce the Digital License Number.Each authorized user has a key that maps to
one of the configuration of all possible permutations. Quadratic Chaining is a method where iteration over(i+i2)Kmod256,where K is the license number,yields64places to insert watermark.Hash Table is another approach to do the same.
Watermarking of MPEG-4Videos
749
Template
Watermarked signal
(a)(b)
Fig.1.a)Schematic Diagram for preparation of watermark signal.b)Outline for
a watermark embedder.The embedder mimics the MPEG-4decoder
3.3Adaptive Local Gain Methodology
The adaptive local gain methodology improves the performance of the water-mark.For relatively smooth regions of the video,where even a small amount of distortion may be visible,the local gain control reduces the watermark em-bedding power to minimize the watermark perceptibility.For relatively busy or textured regions of the image,the local gain control increases the embedding power for improved robustness.The local gain method uses a local activity mea-sure to adjust the watermark power on a block by block basis,which is obtained directly from DCT coefficients for intra-blocks and predicted using motion vector information for predicted blocks.The gain model outputs a local gain L(x,y). The watermark coefficients are then weighted by L(x,y)to produce the water-mark signal that will be embedded into video:
W∗(x,y)=αL(x,y)W(x,y)(3) where W*is the watermark that will be embedded,αis user selected global
gain and W is the watermark signal prior to gain adjustment.For each VOP, local gain weights are decided based on the estimated activity in the VOP.For Intracoded-VOP’s the L(x,y)is
L(x,y)=DCT(f(x),g(y))2ΣDCT(i,j)2(4) where f(x)and g(y)map the(x,y)th pixel to(f(x),g(y))DCT coefficient.For predicted VOPs we used the same formula but instead of Total Energy being ΣDCT(i,j)2it is taken to be
A i=N1
N
A A+
N2
N
A B+
N3
N
A C+
N4
N
A D(5)
where N1,N2,N3and N4are the number of pixels that moved from Area A,B,C,D to the block currently being considered.The sketch for local adaptive gain is shown in Fig.2(a).
750Abhinav Gupta and Phalguni
Gupta
(a)
(b)(c)
Fig.2.a)The sketch for Local Adaptive Gain Method.b)The Original watermark signal that is to be inserted.c)The final watermark signal after
tiling
(a)Video I(900
Kb/sec)(b)Video II(64Kb/sec)
Fig.3.a)Watermarked Video of bitrates 900Kb/sec and payload bit-rate of 1Kb/frame.b)Watermarked Video of bitrates 64Kb/sec and payload bit-rate of 1Kb/frame
4Results
The algorithm has been tried with the two videos (one having lot of texture compared to other)taken in IIT Kanpur as there is no standard video dataset for watermarking approaches.The videos were compressed at different bit rates (64-900Kb/sec)and the watermark signal bit rates were varied from 1kb/frame to 3Kb/frame.The frames were 192x 192in dimensions and a watermark of 32x 32was inserted.The preparation of watermark signal has been shown in figures.2(b),2(c).The results of watermark embedding at 1Kb/frame and 900kb/s and 64kb/s videos are shown in Fig.3.Figure 4shows the performance of local adaptive gain module.

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。