The Dynamic Keyframe Selection

Published Date: 02 Nov 2017

Abstract

A s use of Internet grows day by day for sharing huge vol-

ume of data among each other,it is necessary to protect

this data and its originality. This data contains form

of multimedia like photographs, digital music, or digi-

tal video. These contents can be protected from illegal

access and redistribution by hiding authnticational in-

formation.This process is called as watermarking. This

will assure to content providers that their contents can

not be shared illegally.In this paper we propose digital

watermarking technique to secure multimedia contents.

This technique uses Wavelet based Contourlet Trans-

form (WBCT) along with Non-negative Matrix factoriza-

tion (NMF).WBCT eliminates drawbacks of Contourlet Transform which fails to identify smooth curves in im- ages. NMF is used to address problem of negative com- ponents present in Single Value Decomposition (SVD).

Key terms

Watermarking, Contoulet Transform, Wavelet Based

Contourlet Transform, Non-negative Matrix Factoriza-

tion.

1. Introduction

Because of the rapid sucess of network technology, hu- mans can arbitrarily and easily access or distribute any multimedia data from networks.The development of multimedia tools increase possibility of making tam- pering and forgery.So it is important issue to verify originality and integrity of data.Digital watermarking is most accepted method to provide copyright protec- tion.Digital watermark contains identification informa- tion about copyright owner,authorized user etc.

Digital watermarking can be divided into two types: visible and invisible. For visible watermarking, the em- bedded watermark can be visually observed. The wa- termark must not detract from the image content itself. The advantage of visible watermarking is that it is easy to recognize the owner of the image without any cal- culation, but its disadvantage is that the embedded wa- termark can also be easily removed or destroyed.invisible watermarking can be classifiedinto two types: robust and fragile watermarks.They are often used in copyright pro- tection to declare rightful ownership. In contrast to im- age authentication, fragile watermarks are adopted and designed to detect any unauthorized modification such as distortion under the slightest changes to the image.

Video watermarking introduces some issues which is not present in image watermarking. Due to large amounts of data and inherent redundancy between frames, video signals are highly suscepectible to pirate at- tacks, including frame averaging, frame dropping, frame swapping, statistical analysis, etc. Applying a fixed image watermark to each frame in the video leads to problems of maintaining statistical and perceptual in- visibility. Furthermore, such an approach is necessar- ily video independent; as the watermark is fixed while the frame changes. Applying independent watermarks to each frame also presents a problem. Regions in each video frame with little or no motion remain the same frame after frame. Motionless regions may be statisti- cally compared or averaged to remove independent wa- termarks [2]. In addition, video watermarking schemes must not use the original video during watermark detec- tion as the video usually is in very large size and it is inconvenient to store it twice.

Although the wavelet transform has been proven to be powerful in image processing applications like water- marking; wavelets are not optimal in capturing the two- dimensional singularities found in images. Therefore, several transforms have been proposed for image signals that have incorporated directionality and multiresolution and hence, could more efficiently capture edges in natu- ral images. The contourlet transform is one of the new geometrical image transforms, which can efficiently rep- resent images containing contours and textures. In the contourlet transform, a Laplacian pyramid is employed in the first stage, while directional filter banks (DFB) are used in the angular decomposition stage [1]. Due to the redundancy of the Laplacian pyramid, the con- tourlet transform has a redundancy factor of 4/3 and hence, it may not be the optimum choice for image cod- ing applications. Also One major drawback of SVD is that the basis vectors may have both positive and neg- ative components and the data are represented as linear combinations of basis vectors of positive and negative co- efficients. In many applications the negative components contradict physical realities.

1

In this paper we first propose a new non-redundant im- age transform, the Wavelet-Based Contourlet Transform (WBCT), with a construction similar to the contourlet transform. Then, we use the non-redundant WBCT in conjunction with an SPIHT-like algorithm [10] to con- struct an embedded image coder. To overcome drawback of SVD NMF is used to obtain frequency region of key video frame.

This paper is organized five sections. The subsequent section explains the important aspects of video water- marking. Section 3 focuses the widespread applications of video watermarking. Section 4 considers the robust- ness aspect by elaborating on the common attacks in video watermarking. The various domains of video wa- termarking are explored and a robust algorithm in each domain is considered in Section 5.

2. Related Work

For watermarking purpose video first gets divided into frames which is then manipulated as still image.Images can be represented in spatial domain and transform do- main. The transform domain image is represented in terms of its frequencies; however, in spatial domain it is represented by pixels. In simple terms transform domain means the image is segmented into multiple frequency bands. To transfer an image to its frequency representa- tion we can use several reversible transform like Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), or Discrete Fourier Transform (DFT). Each of these transforms has its own characteristics and repre- sents the image in different ways.

2.1. DCT Domain Watermarking

Steps in DCT Block Based Watermarking Algorithm:

1. Segment the image into non-overlapping blocks of

8x8.

2. Apply forward DCT to each of these blocks.

3. Apply some block selection criteria.

4. Apply coefficient selection criteria.

5. Embed watermark by modifying the selected coeffi- cients.

6. Apply inverse DCT transform on each block.

Disadvantages:

1. In DCT images are broken into blocks 8x8 or 16x16 or bigger. The problem with these blocks is that when the image is reduced to higher compression ratios, these blocks become visible. This has been termed as the blocking effect.

2. Does not perform efficiently for binary images (fax or pictures of fingerprints) characterized by large periods of constant amplitude (low spatial frequen- cies), followed by brief periods of sharp transitions

2.2. DWT Domain Watermarking

The wavelet transform decomposes the image into three spatial directions, i.e. horizontal, vertical and diagonal. Hence wavelets reflect the anisotropic properties of HVS more precisely. Single level decomposition gives four fre- quency representations of the images. These four repre- sentations are called the LL, LH, HL, HH subbands. One of these frequncy band is selected to insert watermark. Disdavantages:

1. The cost of computing DWT as compared to DCT

may be higher.

2. The use of larger DWT basis functions or wavelet filters produces blurring and ringing noise near edge regions in images or video frames.

3. Longer compression time

2.3. Geometric Transformations

The contourlet transform allow for different number of directions at each scale/resolution to achieve a criti- cal sampling. The Contourlet transform has good ap- proximation properties for smooth 2D functions, finds a direct discrete-space construction, and is therefore computationally efficient. For this purpose, contourlet seems to be an appropriate candidate for image com- pression purpose. The contourlet transform is a new directional transform,which is capable of capturing con- tours and fine details in images.It is realized as a dou- ble iterated filter bank. The Discrete Contourlet Trans- form is also called as Pyramidal Directional Filter Bank (PDFB).Like wavelets, contourlets have a seamless trans- lation between the continuous and the discrete domain via multiresolution framework and iterated filter banks.

The contourlet transform is one of the new geometrical image transforms, which can efficiently represent images containing contours and textures.In the contourlet trans- form, a Laplacian Pyramid is employed for the subband decomposition.Due to the redundancy of the Laplacian

2

pyramid, the contourlet transform has a redundancy fac- tor of 4/3 and hence, it may not be the optimum choice for image coding applications.

3. Programmer’s design

Let j is each level in the wavelet transform. LH, HL, and HH are the traditional three highpass bands. We have 2l directional subbands with Gl , 0 ≤ k < 2l equivalent synthesis filters and the overall downsampling matrices of Sl , 0 ≤ k < 2l defined as:

kThe proposed system implements video watermarking where input video is divided into number of frames.

kS(l) =

 2l−1



a0

k



These frames are given as input to watermarking algo- rithm to apply wavelet-based contourlet transform fol- lowed by non negative matrix factorization.We will get watermarked image with watermark embedded in Y com- poenent. Thses frames will be combined to form final

(" 2

#a0



0 a2

output.Here we will discuss mathematical model for pro- posed system.

0 a2l−1

Then,gl [n − Sl m], 0 ≤ k < 2l , m ∈ Z 2 is a directional

3.1. Mathematical Model

3.1.1. Dynamic Keyframe Selection

k

basis for l2 (Z

filter Gl

k

k2 ) gl is the impulse response of the synthesis

Let K be key used to select frame. It is of 80 bits where

8 bits for each digit. Offset for each digit from Y1,....,Y5

k . Assuming an orthonormal separable wavelet

transform, we will have separable 2-D multiresolution:

is calculated using following functions: If the key is a b V 2

2 2 2

c d e f g h i and j respectively then

Y 1 = (e + f ) + 2(d + g) + 3(c + h) + 4(b + i) + 5(a + j) Y2 = (e+f) +2(c+h) +3(d+g) +4(a+j) +5(b+i)

2where W 2

j = Vj ⊗ Vj , andVj−1 = Vj ⊕ Wj

Y3 = (e+f) +2(a+j) +3(b+i) +4(c+h) +5(d+g)

j is the detail space and orthogonal component

of Vj inVj−1 . The family ψj,n , ψj,n , ψj,n n2 is an or-

2 2 1 2 3

Y4 = (Y1+Y2)-Y3

Y5 = (Y1+Y3)-Y2

The value of above function which is multiple of 300 chosen as frame number.

thonormal basis of Wj . Now, if we apply lj directional levels to the detail multiresolution space W 2 ,we obtain

j

j2lj directional subbands of W 2 :

3.1.2. Extract Y component

W 2 2lj−1

2,(lj)

YUV is a color space typically used as part of a color image.The Y’UV model defines a color space in terms of one luma (Y’) and two chrominance (UV) compo- nents.Y’UV signals are typically created from RGB.

Defining the following constants:

Defining

ηi,(lj)

j = ⊕k=0 Wj,k

 l hm − S( lj) ni Ψi

WR = 0.299

WB = 0.114

WG = 1 − WR − WB = 0.587

j,k,n =

m∈Z 2

g jk

k j,m , i = 1, 2, 3

UM ax = 0.436

n 1,(lj)

2,(lj)

3,(lj) o 2

VM ax = 0.615

Y’UV is computed from RGB as follows:

the family

ηj,k,n , ηj,k,n , ηj,k,n

2,(lj)

Z is a basis for

∈n

Y 0 = WR R + WG G + WB B

U = UM ax B−Y ≈ 0.492(B − Y 0)0

1−WB

V = VM ax R−Y ≈ 0.877(R − Y 0)0

1−WR

3.1.3. Wavelet Based Contourlet Transform

The WBCT consists of two filter bank stages. The first stage provides subband decomposition, which in the case of the WBCT is a wavelet transform, in contrast to the Laplacian pyramid used in contourlets. The second stage of the WBCT is a directional filter bank (DFB), which provides angular decomposition. The first stage is real- ized by separable filter banks, while we implement the second stage using non-separable filter banks.

the subspace Wj,k

3.1.4. Non-Negative Matrix Factorization

Let V be a non-negative matrix of dimension: n x m. NMF algorithm decomposes or factors the matrix into a low rank,sparse and non-negative factors such that the original data can be approximated as

V ≈ W H

W is of dimension n x r which have non-negative ele- ments.H is of dimension r x m.

3

3.2. Dynamic Programming and Serialization

The proposed system follows divide and conquer strategy where aim is divided into subgoals. For watermarking using proposed algorithm modules are divided as follows:

1. Reading video

2. Dividing video into number of frames.

3. Extract Y component from each frame.

4. Apply WBCT on these frames.

5. Perform NMF on low frequency part of frame.

6. Map watermark image to factorized component us- ing function.

3.2.1. Memorization parameters

In computing, memoization is an optimization technique used primarily to speed up computer programs by having function calls avoid repeating the calculation of results for previously processed inputs. In Following table we will discuss various memorization parameters identified for different functions.

Function

Memorization Pa-

rameter

Read video

Format of video, resolu-

tion of video

Dividing video

Number of frames, Offset

of frames

Extract Y com-

ponent

Frame offset

Apply WBCT

Offset of frame to be wa-

termarked.

Perform NMF

Low frequency band of

frame.

Watermark em-

bedding

NMF matrices produced

for watermark image and video frame

Table 1: Memorization Parameters

3.3. Data independence and Data Flow architecture

3.3.1. Data Design

Data independence is the type of data transparency that matters for a centralized DBMS. It refers to the immu- nity of user applications to make changes in the definition and organization of data.For proposed system folowing tables are used to achieve logical data independence.

Field Name

Data Type

Name of Video

String

Type of Video

String

Table 2: Video Table

3.3.2. Data Flow Architecture

For video watermarking system data serialization is shown using data flow diagram and process tree. These structures represent different activities present in system implementation and their flow.

• Data Flow Architecture A data flow diagram (DFD) is a graphical representation of the "flow" of data through an information system, modeling its process aspects.A DFD shows what kinds of information will be input to and output from the system, where the data will come from and go to, and where the data will be stored.

The above diagram give flow of algorithm. Video watermarking algorithm is divided into number of sub-algorithms which perform specific task using in- put. Labels on arrow represent input for next sub- process and output generated by previous function.

• Process Tree Process tree for proposed system con- tains heirarchy of different functions system will exe- cute as process. The diagram reflects such processes and subprocesses present in proposed system.

3.4. Multiplexer Logic

The growing and computational power and programma- bility of the of multi-core architectures provide great prospects for acceleration of image processing and com- puter vision algorithms which can be parallelized.The op- erations performed by image processing algorithms can be computationally expensive due to their manipulat- ing large amount of data. To make a program execute in real-time, the data needs to be processed in parallel and often a great deal of optimization needs to be uti- lized.This high degree of natural parallelism exhibited by most of the image processing algorithms can be easily ex- ploited using SIMD parallel architectures and computing techniques.[9]

Recently, a number of novel and massively-parallel computer architectures have been introduced that promise significant acceleration of applications by using

a large number of compute cores. The super pipelined

4

Figure 2: Process Tree

Figure 1: DFD of Proposed System

processor design approach that pushes the limits of per- formance by increasing the pipeline length has hit the power wall paving the way for multi-core and/or multi- threaded architectures. GP GPUs take advantage of the GPU architectures stream model of computation. The IBM cell processor is another leading commercial pro- cessor based on the stream model of computation.[9]

In the world of desktop computers dominated by x86- based processors, the super pipelined processor design approach that pushes the limits of performance by in- creasing the pipeline length has hit the power wall paving the way for multi-core and/or multithreaded architec- tures to invade this market segment. In the world of mo- bile computers, novel low power design techniques have been adopted in the processor, chipset, and system to maximize the battery life while keeping the performance

at acceptable levels. Also, virtualization and security support are now visible in many product offerings.[9]

• NVIDIA G200 GRAPHICS/High-Performance

Compute

The NVIDIA G200 is a high-performance ar- chitecture specifically aimed at data dominated applications, particularly raster graphics for exam- ple medical imaging, and financial data processing. However, it is also able to provide more general programmability to support nongraphics related, data dependant applications. The memory system is noncoherent and uses small local stores instead of a standard cache style architecture. The G200 does provide some facility for more general parallel programs by providing âĂIJatomic operationâĂİ units. These are used for controlling access to shared data structures that live in the GPUâĂŹs main memory.[8]

• INTEL CORE I7 GENERAL PURPOSE

The Intel Core i7 [2] is a high-performance general-

5

purpose processor in all respects. It attempts to do everything well.The memory system is typical of that found in a generalpurpose multicore machine with just a few cores. It uses a fully coherent mem- ory system and has large standard caches. The co- herence is broadcast based, which is sufficient be- cause of the limited number of cores. These charac- teristics come together to create a chip that is good at a wide variety of applications provided power is not a constraint.[8]

3.5. Turing Machine

a Turing machine can be adapted to simulate the logic of any computer algorithm, and is particularly useful in explaining the functions of a CPU inside a com- puter.Following diagram shows different states of the sys- tem through which it will go during project execution and type of input it accepts output it will produce.

The system will start with reading state where input video is accepted. This video gets divided into number

of frames in second state. In next step Y component of frame is extracted in which watermark embedding will be done. Y component is given to apply WBCT. Low fre- quency part of this component is factorized using NMF and mapped with factorized watermark image.Finally all frmaes gets combined to form watermarked video.

4. Results and Discussion

In the experiments, we use the video of .avi format as input for watermarking.Image of .PNG format is used as watermark.The experiment is performed by taking scal- ing factor alpha as 0.05 to 0.5 in the steps of 0.01.From the result one can decide there are no perceptibly vi- sual degradations on the watermarked image. Following are results: Video frame extraction Result tables show- ing data regarding the algorithmic efficiency parameters and comparative discussion related to efficiency of the algorithm proposed.

5. Conclusion

In this paper, we propose a nonblind hybrid video wa- termarking scheme for video copyright protection using wavelet based contourlet transform and Nonnegative ma- trix factorization. The embedding is performed by adap- tive changing the NMF factors of low frequency subband of Y component of each video frame with respect to NMFs of watermark image.

As future work we will apply NMF on Y component of low frequency bands with sparseness constrains which uses a nonlinear projection operator to achieve the pre- cise control of the sparseness by adding sparseness con- straints in all basis vectors.