Robust Video Watermarking Using a Hybrid DCT-DWT Approach

Jie Sang | Qi Liu | Chun-Lin Song

Abstract—In order to solve the limitations of the digital video watermarking algorithm,this paper proposes a new robust video watermarking algorithm using combining discrete cosine transform (DCT) and discrete wavelet transform (DWT) techniques.First of all,the video frames are randomly selected and then the DCT algorithm is applied to the selected video frames.After that,the first column of the selected video frames is scrambled using the Arnold algorithm.Furthermore,every column with 4 direct current (DC) coefficients is reshaped and transformed into four different sub-bands using the DWT technique.Next,the watermark is embedded into the approximation (LL)sub-band.The proposed algorithm is easy to carry out because it provides random frames with no special requirements for video frames.The experiment results indicate that this algorithm can resist against different kinds of watermarking attacks,such as the Gaussian filter attack and sharpen attack.In addition,it also illustrates that the proposed algorithm has a better result than some other watermarking algorithms.

1.Introduction

With the rapid development of the Internet and digital technologies,the distribution of digital products on the Internet becomes common business.It can benefit our lives,for example,digital products such as electric books,digital images,movies,and music,and also can be easily accessed in anytime,anywhere,and ways which are much more convenient than ever before.However,it also causes some challenges[1].Under these circumstances,digital products are becoming more easily duplicated,reproduced,distributed,or exhibited without the proper authorization of the product owner.In other words,the copyright of the digital products’ owner is severely threatened.It has been suggested that the digital watermarking technology can be used as an effective method to solve copyright protection problems and play an important role in digital products tracking,monitoring,authentication,etc[2].

A complete watermarking process should have three important functions,including watermark embedding,watermark extraction,and watermark detection.It usually embeds invisible copyright watermarks into the host data and does an additional encryption.These invisible copyright watermarks are certification,symbols,digital signatures,and so on.The suitability of a digital watermarking algorithm in copyright protection is determined by many factors,one more important of which is its robustness to different attacks.Therefore,improving algorithm robustness to different attacks has been one of the main objectives of digital watermarking research[3].

Different types of digital contents,such as still images,videos,or documents,often require different watermarking algorithms[4].This paper concentrates on solving the issues in digital video watermarking.Compared with the ordinary digital images,the data volume of the digital video is much larger.Therefore,high-efficiency and high-robustness are the main issues on embedding and extracting algorithms towards video watermarking.Generally speaking,there are three types of video watermark embedding strategies,namely the non-compressed original video frame embedding technique,video code embedding technique,and compressed video stream embedding technique.This paper applies the first watermark embedding strategy for inserting watermark signals into different video frames.

There are three criteria to measure the performance of each watermarking algorithm,which are named fidelity,data-payload,and embedding effectiveness.This paper makes a trade-off between these criteria,and pays more attention to the robustness of the video watermarking algorithm.This is aimed to preserve the proof of ownership of the videos.We do this by combining the two most widely used transform algorithms,discrete cosine transform (DCT) and discrete wavelet transform (DWT).They are used together to produce a better result on resisting different types of watermark attacks.The watermark embedding performs more effectively in the transform domain rather than that in the spatial domain,and we hypothesize that it is the main reason for the achieved improvement.This paper improves the robustness of video watermarking.The time domain of the video is transformed into the frequency domain,which is extracted and re-integrated through the special matrix transformation.In addition,direct current (DC) and approximation (LL) domains are selected for watermark embedding,which greatly improves the timeliness and robustness of watermarking.In the face of different physical and removal attacks,this algorithm shows more robustness compared with other algorithms.In addition,the main contributions of this hybrid algorithm are shown as below:

1) A robust and efficient hybrid algorithm for video watermarking is described;

2) A unique matrix transformation for dimension reduction and data reduction is described;

3) The robustness of embedding and extracting is improved obviously;

4) The 64 random selected video frames are watermarked,which is more robust to resist the physical attack;

5) DCT and DWT are used jointly,which combines the advantages of both;

6) The complex transforming process makes the hybrid algorithm safer,because cracking such transformation is difficult.

The rest of this paper is as below:Section 2 provides a brief description of DCT and DWT techniques.The details of the proposed watermarking technique are described in Section 3.The experiment results and the analysis under different attacks are shown in Section 4.And the final conclusion is given in Section 5.

2.Transformation Techniques

In the section,we provide a brief description of DCT and DWT techniques.Their applications used in the digital watermarking and non-compressed situation are also introduced.

2.1.DCT AIgorithm

DCT is an orthogonal transformation,which plays an important role in digital image and video processing.DCT is a transform coding technique which is widely used in multimedia compression and digital watermarking.Supposing that a video framef(m,n) is transformed by the DCT algorithm,the video frame can be represented as a 2-dimensional (2D) matrix with the size ofm×n.The technique first separates the 2D matrix into several non-overlapping and identical-sized blocks.The size of the block is decided by the requirements of the video frame’s quality,its compression ratio,and computational complexity[5].The size is typically 8×8 pixels.The next step applies the DCT algorithm to each block,which is shown in Fig.1.We can notice that the DCT coefficient at the top left site is the DC coefficient,which represents the average value.The rest of the remaining coefficients are alternating current (AC)coefficients,in which the highest horizontal frequency is located at the right side and the highest vertical frequency is placed at the bottom.The inverse DCT (IDCT) is also applied by reconstructing and decompressing these coefficients.

Mathematically,the process of calculating DCT is denoted asF(s,t),and a video framef(m,n) is shown in (1):

whereMandNrepresent the width and height of the video frame in pixels,respectively.The values ofc(s)andc(t) are calculated as

The inverse transform of IDCT,denoted asF(m,n),is calculated by using

Although the DCT algorithm has a lot of advantages in image and video watermarking,however,it also has some shortcomings such as “blocking artifacts” especially at the low bit rates.To improve the shortcomings,in this technique,we use DC coefficients only and apply the DWT algorithm to embed the watermark content in the further steps.

2.2.DWT AIgorithm

DWT is an information analysis theory and a signal of the space and time scale analysis method,which is quite popular in recent years.It has multiple scales in the space and frequency,whose decomposition of the image can be carried out continuously from low resolution to high resolution.Furthermore,the DWT algorithm is widely applied in the watermarking field.Researchers have produced numerous innovations and efficient joint algorithms related to the DWT algorithm[6].

This transform is expressed in various wavelets which are composed by different small waves in a limited period with different frequencies.A wavelet series is a square integrable function by a certain orthogonal series generated by the wavelet transformation.In addition,DWT can decompose the original signal into the wavelet transform coefficients that contain the position information.Finally,the wavelets could apply the inverse DWT(IDWT) algorithm to reconstruct the signal completely.

The DWT technique divides the original content into four sub-bands,including the LL sub-band,vertical (LH)sub-band,horizontal (HL) sub-band,and diagonal (HH) sub-band.The LL sub-band is composed by the low-pass filter in both horizontal and vertical directions,which roughly describes the video frame or image content,and the LL components contain the most of the original information.The LH and HL sub-bands are formed by low-pass and high-pass filters filtering in different directions.The LH sub-band contains the vertical information of the horizontal edges,while the HL sub-band contains the horizontal information of the vertical edges.Finally,the HH sub-band is formed by the high-pass filter in the same directions,which contains the high-frequency components.The details of the DWT decomposition process are shown in Figs.2 and 3.

Fig.2.Decomposition process of DWT.

Fig.3.One example of the decomposition step.

Fig.2 shows an example of the decomposition process when the input signal isX[m,n].The two-step process to produce the decomposition signal into its LL component is shown in (5) and (6).

whereg[k] is the low-pass filter;V1,L[m,n] represents the single-level low component of the original component.

The DWT algorithm is widely used in digital watermarking.In such application,DWT creates the masking effect which makes a trade-off between the properties of transparency,capacity,and robustness.In most of the cases,the LL sub-band is chosen as the watermarked region because embedding a watermark in the low frequency subband may increase robustness significantly.

At present,there are many watermark algorithms proposed by using DCT,DWT,or their combination.However,in these achievements,most of them focus on the digital images[7]and only a few of them are applied on digital videos[8].In most of the digital video watermarking process,the first step is to select the key frames of digital videos,after that,the digital watermarking algorithm is applied to such frames to protect the copyright and trace the pirate[9].Such an example was given in 2012 by Deyet al.[10]who proposed a hybrid algorithm using the DCT-DWTsingle value decomposition (DCT-DWT-SVD) algorithm.In the experiment part,there is a comparison experiment between the proposed algorithm and DCT-DWT-SVD algorithm.This paper proposes a new robust hybrid watermarking algorithm based on the DCT-DWT hybrid algorithm,which tries to resist against various different watermark attacks.

2.3.Hybrid DWT-DCT AIgorithm

With the development of watermarking algorithms,the combination of the DWT and DCT algorithms is widely used in watermarking.Traditionally,DWT is applied to the entire image and allows image decomposition into separate sub-bands for the purpose of showing multiply resolution representation.The LL sub-band,for example,can be selected to achieve a robust characteristic,whereas the HH sub-band can be used to preserve the finer detailed information.On the other hand,DCT is applied to square-sized blocks of the image as opposed to the entire image.This allows a high compression ratio to be achieved when DCT is used for compression purposes.For watermarking purposes,the watermarks can be embedded onto the DC coefficient to achieve robustness.Meanwhile,the watermarks can be embedded onto either low frequency or high frequency AC coefficients to preserve the detailed information.The jointed DCT-DWT algorithm takes full advantages of their strength and is widely applied in image watermarking[11],[12].However,there are only very few hybrid DCT-DWT watermarking algorithms for digital videos at present.

3.Proposed Hybrid AIgorithm Based on DCT and DWT

This section introduces the completed process of the proposed DCT-DWT digital video watermarking algorithm,including watermark insertion and extraction.This proposed hybrid algorithm focuses on the robustness of video watermarking,which chooses the most robust domains named DC coefficients and LL sub-bands.The unique matrix transformation helps the 3-dimensional (3D) video frames into the 2D space.Unlike image watermarking,this algorithm embeds the watermark into many video frames in order to resist different signal and physical attacks.

3.1.Watermark Insertion

Generally speaking,the hybrid digital video watermarking algorithm we proposed in this section embeds the watermark contents into the random-selected 64 video frames.To improve the performance and robustness of the proposed algorithm,we decide to use 32×32 DC coefficients in each selected frame[13]and use LL sub-bands to finish the insertion process.Both of the DC coefficient and LL sub-band belong to the low frequency domain,therefore it contains the most information of the original video frames.The details of the insertion process can be seen in Fig.4.Furthermore,due to the properties of DCT and DWT,the hybrid algorithm contains a good robustness characteristic and the details can be found in the next section.

As shown in Fig.4,the insertion process contains several steps and the details are described as below:

1) To start with,64 frames are randomly selected from the original digital video.The luminance component of each frame is extracted,which is denoted asY.

2)Yis separated into 64 non-overlapping blocks and then 2D-DCT is applied.

3) The 32×32 DC coefficients are extracted from each of the blocks and then re-grouped to create a new 2D matrix,which is referred as Group_DC.The new matrix contains 4 different rows; each row is composed of 16 frames and each frame has 32×32 DC coefficients.The process is shown in Fig.5.

As illustrated in Fig.5,the 64 frames of 32×32 DC coefficients are transformed into a 4×16384 matrix.The matrix is constructed in such a way that each frame is rearranged by the first-row with 1×1024 vectors.The first four of these vectors (from the first 4 frames) are arranged column by column to make the left-most 4×1024 elements of the matrix.The process is repeated for the next four vectors and the result is appended to the right of the previous result,and so on.The resulting output is a 4×16384 matrix called Group_DC.

4) After that,there are four different DC coefficients randomly selected from different rows to create a 2×2 matrix for further processing.The complex matrix transform is aimed to prevent the watermark content from being attacked because there are so many DC coefficients that could be selected as the potential embedding points.

Fig.4.Watermark insertion process.

Fig.5.Detailed process of the matrix transformation.

5) Then,we apply the Arnold transform[14]for the watermark content to get another additional protection level on the watermark information.

6) At each column of Group_DC,we apply 2D-DWT to get LL,LH,HL,and HH sub-bands.Next,we extract LL coefficients and group them to become a new matrix which is named as Group_LL.

7) The watermark content is transferred into a proposed layout according to the length of Group_LL and the spread spectrum technique is applied on the watermark signal.

8) Finally,the watermark content is embedded into the LL sub-band by using

whereαis the embedding strength and Group_LL_watermarked is the watermarked content.

9) The IDWT algorithm using the modified DWT coefficient is applied to produce watermarked_Group_DC'.

10) The inverse-Arnold transform algorithm is applied on watermarked_Group_DC'.

11) Finally,the IDCT algorithm is applied to watermarked_Group_DC' to restructure each luminance component of the video frame to get the watermarked video.

3.2.Watermark Extraction

The proposed watermark extraction process requires the information of the randomly-selected frames.The specific process is shown in Fig.6.

Fig.6.Watermark extraction process.

1) First of all,the 64 watermarked video frames are selected and then the luminance component in each frame is extracted,which is denoted asY'.

2) SeparateY'into several non-overlapped 8×8 blocks and then 2D-DCT is applied.

3) Extract 32×32 DC coefficients in each selected frame,re-matrix and group them into a new Group_DC matrix.

4) Apply the Arnold transform to get the new matrix.

5) Each column of the new matrix is transformed to a 2×2 matrix and 2D-DWT is applied to get LL,LH,HL,and HH sub-bands.All LL coefficients are extracted and grouped into a new matrix which is named as Group_LL_extract.

6) Finally,we extract the watermark content using

7) Reshape the watermark content to video frame content and apply the inverse-spread spectrum technique on the watermark content.Then the extracted watermark is gotten.

4.Experiment ResuIts and AnaIysis

In this paper,we apply the DCT-DWT algorithm to produce a novel video watermark scheme.A comparison experiment between the proposed algorithm and DCT-DWT-SVD algorithm is shown here.The DCT-DWT-SVD algorithm uses DWT first to choose the HH sub-band,then applies the DCT algorithm to the HH sub-band,and uses SVD to embed the watermark signal.In this section,we test the robustness of the proposed hybrid algorithm.Tempete.yuv and Football.yuv[15],[16]as the standard video templates are applied in this section.Tempete.yuv shows the leaves swing and Football.yuv shows the football sports.The dimension of the host video frame is 288×352 pixels and the watermark image is sized by 32×32.The maximum embedding capacity is 101376 coefficients.The embedding strengthɑis 1.0.Fig.7 shows the first frames of the host video and watermarked content.In this section,the peak signal to noise ratio (PSNR) and normalized correlation (NC) are used to describe the experiment result.PSNR is always used to describe the similarity between the original and watermarked images,while NC is always used to compare the original and extracted watermarks[17].

Fig.7.Original and watermarked video frames:(a) and (b) are the original video frames; (c) and (d) are the watermarked video frames.

4.1.InvisibIe QuaIity and Watermark Integrity

To measure the quality between the original video frame and the watermarked video frame,PSNR is applied,as

where MSE represents the mean square error and is presented as below:

If PSNR is greater than 30,it means the transparency of the watermarked video frame keeps the similarity as the host video frame.Therefore,Table 1 shows the PSNR values between the original video frame and the watermarked video frame.Fig.7 shows the images of them.As can be seen in Table 1,the proposed algorithm gets larger PSNR values than the DCT-DWTSVD algorithm.In addition,we also find that the watermark signal is successfully embedded into the original video frame and the properties of transparency are not affected.

In addition,the NC value is always used to compare the similarity between the extracted watermark signal and the original watermark signal.The range of the NC value is between 0.0 and 1.0.The formula of NC is shown as

Table 1:PSNR between the original and watermarked video frames

If the NC value closes to 1.0,it represents that the extracted watermark content is similar to the original watermark.The comparison of the original watermark content and the extracted watermark content is shown in Fig.8.Table 2 compares the NC values of the proposed algorithm with that of the other one.As that can be seen in Table 2,we find the quality of the extracted watermarks obtained from the proposed algorithm is better than that of the DCT-DWT-SVD algorithm.

Fig.8.Watermark content:(a) original and (b) extracted watermarks.

Table 2:NC between original and extracted watermarks

4.2.AIgorithm QuaIity under Different Attacks

This part tests the robustness of the proposed algorithm under different watermarking attacks.There are six watermarked attacks proposed in the experiment,they are the Poisson attack,salt-pepper filter attack,Gaussian filter attack,cutting frames attack,sharpen attack,and enlarge and revert attack.The severity of these attacks can be adjusted by modifying their corresponding parameter values.The watermarked video frame in Fig.7 (b) is suffered from these attacks and the corresponding extracted watermarks are shown in Fig.9.The comparison of NC values between the proposed and original algorithms is shown in Table 3.

Fig.9.Extracted watermarks after (a) Poisson filter attack,(b) salt-pepper filter (0.02) attack,(c) Gaussian filter (0.05)attack,(d) cutting frames (5) attack,(e) cutting frames (10) attack,(f) sharpening (linear) attack,and (g) enlarge and revert (2) attack.

Table 3:NC values under different attacks

As shown in Fig.9,the watermark is extracted successfully after different watermark attacks,the results in Tables 1,2,and 3 are also provided the similar conclusion.In addition,we find that most of the NC values are over 0.9,it indicates the proposed algorithm can overcome several different attacks.In addition,from the inspection of Table 3,we find that the proposed technique is significantly more robust to attacks than the compared algorithm.This DCT-DWT-SVD algorithm could not resist against various watermark attacks due to the use of the HH subb and which makes it relatively easy to be attacked.We can also argue that the robustness of our hybrid algorithm is greater than that of another algorithm under different attacks.

5.ConcIusion

For protecting the digital video copyright,this paper proposed a novel robust video watermarking algorithm based on the DCT-DWT technique.It is easy to apply in practice by using the video luminance component.Besides that,we used the DCT technique to transform the spatial video frame into the frequency domain.The watermark was spread before being embedded into the different selected video frames.Finally,we used the DWT algorithm to finish the whole embedding process.The unique matrix transformation made the selected video frames into a 4×16384 matrix,which made the embedding and extracting processes more robust and efficient.The embedding strength was 1.0 in the experiment.When the embedding strength was stronger,the robustness was greater and the visual experience was worse.After several experiments,the embedding strength of this algorithm was more suitable from 1.0 to 1.2.In particular,this algorithm showed high robustness when embedding multiple watermarks.In addition,the embedding process did not face the whole video,but only 64 key frames.The result showed that the proposed algorithm had strong ability to resist different watermark attacks.The extracted watermarks under different attacks were all recognized and most of the NC values were over 0.9.Therefore,the proposed algorithm proves that it is suitable for solving the unauthorized problems in digital videos.The most contribution of this paper was the DCT-DWT hybrid algorithm,which proposed a unique matrix transformation and greatly improved the robustness of the video algorithm.In addition,the hybrid algorithm helped the watermark separate and embed into many different video frames,which also highly improved the robustness in the physical side.