Figure 1. Watermarking analogies.

ABSTRACT

In this paper, a content dependent visible watermarking scheme for video is proposed. We calculate the unsuitability contrary [1], to the focus region that catches user’s attention the most and embed a visible watermark to the region with the highest unsuitability. The watermark embedding is performed shot by shot, so that the watermark’s movement will not be too annoying.

KEY WORDS

Visible video watermarking, user attention model, focus detection framework, invisible video &image watermarking.

1. Introduction

Digital watermarking is defined as a

Process of embedding data, called a watermark, into a multimedia objects such that the embedded watermark can be detected or extracted later to make an assertion about the objects. The multimedia objects can be text, image, audio, video and their compositions.

The end of the previous millennium has seen the transition from the analog to the digital world. Nowadays, audio CDs, Internet and DVDs are more and

More widespread. However film and music content owners are still reluctant to release digital content. This is mainly due to the fact that if digital content is left unprotected, it can be copied rapidly, perfectly, at large scale, without any limitation on the number of copies and distributed easily e.g. via Internet. Protection of digital content has relied for a long time on encryption

But it appeared that encryption alone is not sufficient enough to protect digital data all along its lifetime. Sooner or later, digital content has to be decrypted in order to be eventually presented to the human consumer. At this very moment, the protection offered by encryption no longer exists and a user may duplicate or manipulate it.

Digital watermarking techniques can be divided into various categories basing on the applying domain (spatial or frequency domain) and the type of embedding documents (image or video). According to

Human perceptivity, digital watermarks can be roughly divided into two different types, visible and invisible. Visible watermark is a second transparent pattern or image overlaid with the primary (host) image. The embedded watermark appears visible to a casual viewer undergoes a careful inspection. The invisible watermark is embedded in such a way that alternations made to the pixel value are perceptually not noticeable and can be recovered by using an appropriate decoding mechanism.

(I) The watermarking process should be applied automatically with little human Intervention and labor.

(II) A visible watermark should be obvious in both color and monochrome images.

(III) The watermark should be visible yet must not significantly obscure the image details beneath it.

(IV) Synchronization is a problem that cannot be ignored in most video watermarking applications, even in the absence of a malicious attacker. In applications such as secure digital television and broadcast monitoring, the watermark detector may be expected to detect the watermark starting from any arbitrary temporal location within the video signal (as opposed to starting detection from the “beginning” of the video signal), which is known as initial synchronization

(V) The watermark should be spread in a large or important area of the image in order to prevent its deletion by clipping.

Figure 1. Watermarking analogies.

A few visible image-watermarking schemes have been proposed so far [2,3]. The authors in [3] defined a measure equation to modify block DCT coefficients of an image basing on a mathematical model, developed by exploiting the texture sensitivity of the human visual system (HVS).

However, in video watermarking, the adaptation of

HVS is more difficult compared to the image data.

There are two main reasons in video watermarking that

Makes the problem different. The first one is the Necessity of a guaranteed robustness against new type of attacks, such as frame dropping, frame averaging, and collusion, which have no counterparts in image watermarking. The second one is due to the temporal imperceptibility of the watermark, which is relatively more difficult problem, due to 3-D characteristics of video. In order to provide an imperceptible watermark, the watermarking procedure should also take the variations in the temporal direction into account.

All of the prescribed visible watermarking scheme is trying to do the same thing: modeling the content property from the low-level features (edge, background texture, or intensity) as a clue to adapt the embedded

Watermarks to the hosting contents. We denote this kind of embedding schemes as content-dependent watermarking. Some of the video watermarking schemes still treat video data as a series of pictures and apply some image-watermarking scheme to each Frame of the video as another solution [4], [5] And [6]. This approach may not good enough for video watermarking because the most significant properties of video, temporal correlation and motion information are not taken into account.

We believe that the content dependent watermarking could benefit from some research results in the multimedia content analysis field. By taking advantage of computational attention models, we have more objective references for choosing the locations to insert watermarks.

2. Content-dependent Visible Video

Watermarking

In this section, we describe the proposed content dependent video watermarking scheme. Firstly, we select the target location for embedding watermarks from the candidate non-focus regions. Secondly, we

Calculate the watermark’s scaling coefficient according to the background attributes; and thirdly, we embed the watermark as a transparent mask onto each frame. At last, we perform the watermarking process on each video shot unit to get the final result.

2.1 Selection of Watermark Location

Focus point detection results

Focus point detection results

the focus point indicates the most attractive region

Where users eager to see. On the contrary, for those regions far away from the focus point are good candidate positions to embed visible watermarks. Besides, in order to decrease the perceptual distortion after watermarking, the regions of lower intensity and higher texture are the most suitable ones. Thus, we define a weighting function to measure each macro block’s unsuitability

Here mn is the size of embedded watermark in a macro block,), (y x denotes the coordinate of each macro block and), (f f y x represents the macro block where the focus point resided in.), (y x fI and), (y x f ñ are functions to measure the strength of local intensity and the degree of texture complexity, respectively. The meterage of intensity is performed on the luminance channel of the YUV color space, in which ij I is the average intensity value of the macroblock), (j i. For characterizing textures, we count the percentage of zero quantized DCT coefficients of each macroblock, known as the ñ-value [8]. Both of the luminance intensity, ij I , and ñ-values, ij ñ , can be directly extracted and Calculated from the MPEG bit streams, in the Macroblock level.

Each macro block’s unsuitability, in the current video frame, is calculated by equation (1). The one with the highest unsuitability is the selected location for embedding the watermarks.

2.2 Watermark Embedding Protocol

Our technique for generating the key schedule at the watermark embedder is illustrated in Figure, consisting of four principal components: Watermark embedder, temporal redundancy control, feature extractor, and key generator.

Watermark embedding protocol for temporal synchronization

Watermark Embedder: The watermark embedder accepts as inputs a frame of the original (unwatermarked) video, Embedding key, and any other side information necessary to embed the watermark into the current frame. The embedding key is supplied from the temporal redundancy control component. Any watermark embedding technique suitable under the assumptions described can be used.

Temporal Redundancy Control: This is the primary means of controlling the amount of temporal redundancy in the key schedule. Temporal redundancy is added to the key schedule by using a single key to watermark multiple frames of the video. Three parameters are used to define the key schedule: The master-embedding key K_E is used as the initial key to

Generate the key schedule. The period, a, is the number of frames watermarked before the key generator is reset. The repeat, ß, is the number of consecutive frames that are watermarked with the same key before a new key is generated The functionality of the temporal redundancy control is shown below:

Water Mark Embedding Procedure: (WEP)

1. Set a=0, b=0, K= K_E , reset the key generator state.

2. Read a frame of the input video.

3. Send the current key K to the watermark embedder to watermark the current frame.

4. Increment a by 1. If a<a, continue to step 5. Otherwise, go to step 1.

5. Increment b by 1. If b=ß, continue to step 6. Otherwise, go to step 2.

6. Use the feature extractor and key generator (both described below) to create a new key K for subsequent frames, set

b=0, and go to step 2.

Figure illustrates an example key schedule for the case of ?=8 and ?=2, with frames depicted in the same shades indicating the same key is used to watermark those frames. Every two consecutive frames are watermarked with the same key. The watermark key is reset to K_E every 8 frames, so frames 0, 1, 8, and 9 are watermarked with key K_E . The key used to watermark frames 2 and 3 is not necessarily identical to the key

used to watermark frames 10 and 11. (The name “period” for a is somewhat a misnomer because the key schedule is not periodic in general.)

Example key schedule

The period parameter ?controls the degree of global redundancy in the key schedule, as it determines how frequently the watermark key generator is reset and K_E is used for embedding the watermark. The frames that are embedded with key K_E are known as resynchronization frames because the detector can search for KE when it is lost. These frames also provide

Synchronization points when the detector performs initial synchronization. Increasing a decreases the frequency when K_E is embedded, thus reducing the temporal redundancy in the watermark. A key schedule generated by ?-1 corresponds to embedding with key K_E for every frame, or a time-invariant key watermark.

The repeat parameter ? controls the degree of local redundancy in the key schedule. A high repeat factor corresponds to increased redundancy and decreased key rate (number of distinct keys used for watermarking embedding per unit time.) Increased local redundancy is beneficial for resisting temporal attacks, such as frame dropping, transposition, and averaging. As the experiments will show, a key schedule having minimal local redundancy (large?,?-1) has similar behavior to an independent key watermark and can be desynchronized by frame averaging and deletion attack.

As mentioned Earlier 2, the embedding key is just one parameter used to determine the embedded watermark, and frames watermarked by the same key do not necessarily imply that the embedded watermark signals in those frames are identical. One example is the technique in [7] , which creates image-dependent watermark signals using a time-invariant key.

Feature Extraction: Feature extraction allows the key schedule to be video dependent. It is invoked when a new watermark key is needed (step 6 of the WEP shown above) and is inactive during other times. The feature extractor examines the watermarked image and outputs a vector of features that is made available to the key generator. For robust video watermarking, the features should change in value only when significant alterations are made to the watermarked frame. In our technique, it is possible to destroy temporal synchronization by performing spatial attacks on Video frames because the key schedule is video-dependent. Ideally, the features are sufficiently robust so that such an attack is successful only when the attacked video no longer has any value in the application. The features should also resist (or be invariant to) spatial synchronization attacks and be computationally efficient. In particular, if real-time embedding or detection is required by an application, the feature extraction process must also be in real-time. The features themselves should be kept secret and known only to the embedder and detector. For security, the feature values can involve K_E or an auxiliary key.

Key Generator: The key generator is used to create new watermarking keys in the key schedule. The key generator, which is in general a state machine, is involved in steps 1 and 6 in the Water Mark Embedding Procedure. During step 1, the state of the key generator is reset to a predefined initial state. This is necessary so the watermark detector (which uses an identical key generator) will be able to recover the key sequence. In step 6, the key generator creates a new key using its current state information and input from the feature ex tractor. The state of the key generator can also change in step 6. During other steps of the WEP, the key generator is inactive.

One type of key generator is a finite state machine (FSM). One state in the FSM is identified as the initial state. Each state has an associated key transition function, or a description of how to generate the next key from the current key. Key transition functions can take as inputs the current key, the master key KE, the feature vector supplied by the feature extractor, and other side-information (for example, message payload). State transitions are (in general) dependent on the feature values, however the key generators for time-invariant key and periodic key watermarks can be implemented as trivial FSMs that do not depend on features.

Figure shows an example FSM key generator that uses three features (X, Y, and Z) from the feature extractor. The initial state is state zero. The key transition functions are shown with each state, and with the exception of states 0 and 2, involve the current key value and the value of one of the features (c1, c2, and c3 are constants.) In state 2, the next key is generated by performing binary negation on the current key. The feature (X, Y, or Z) that has greatest value determines the state transitions.

An example key generator using a finite state machine

Like the feature extractor, the structure of the key generator is a shared secret between the embedder and detector. The key generator can also be dependent on KE or an auxiliary key, such as by using key-dependent key transition functions or key-dependent state machines. The flexibility of the key generator and feature extraction components provides a great deal of generality that can be exploited by the application or the degree of security needed. In this paper, we chose a simple set of features and a trivial FSM key generator to implement the protocol and evaluate its performance while varying the temporal redundancy parameters (?,?)

Computational Cost of the Embedding Protocol: The computational costs for watermark embedding can be critical in realtime implementations. By examination of the WEP, the worst-case computational cost is on the order of (WE + FE + KG) per frame, where WE is the computational cost for watermark embedding, FE is the computational cost for feature extraction, and KG is the computational cost for key generation.

2.3 Video Shot Based Watermarking

Since the embedding process is performed on frame level, the selection result may be different from time to time. That implies the embedded watermark may look like another unstable moving object in the video, which is quite annoying. For resolving this shortage, we change our policy to calculate the watermarking location only once in a video shot. The watermark appears in the same location during each shot and only moves when video shot-change occurs. The shot-change detection is realized by using the measurement of minimal square error of pixel values between two adjacent video frames. More advanced shot detection

algorithms can be applied to get better performance and some of them have been discussed in [9].

3. Conclusion

The watermark used for embedding is

approximately quarter of the video frame size, which is large enough for preventing from deletion by clipping and achieving the goal of copyright assertion.. The watermark is prominent in the non-focus regions and can easily be identified by users. Visible video watermarking scheme proposed in [10] generate a randomized location shifting for watermark masks to enhance the security issue, and hence increase the difficulty for attackers to remove the watermark. In our approach, watermark differs in each shot both in its

Location and strength. The appearance of watermark cannot be predicted in advance due to its content-dependent nature, and therefore, can hardly be removed by attackers, either. At last, the proposed watermarking can be done fully automatic, i.e. it is labor saving and without user’s intervention. This goal is not easily to be achieved especially for video watermarking where keeping good perceptual quality is a must. Currently, many research efforts have been devoted to multimedia content analysis, trying to break the barriers between the low-level features and high-level semantics. And we believe that by exploiting more content features with

higher semantic meanings, approaching a better digital watermarking scenario could be possible.

4. Reference

[1] Chia-Chiang Ho, Wen-Huang Cheng,

Ting-Jian Pan, and Ja-Ling Wu, “A

User-Attention Based Focus Detection

Framework and Its Applications”, the 4th

IEEE Pacific-Rim Conference on Multimedia,

2003

[2] Yeung M.M., et al., "Digital Watermarking for High- Quality Imaging", Proc. of IEEE First Workshop on Multimedia Signal Processing, Princeton, NJ,

pp. 357-362, June 1997.

[3] F. Mintzer, G. Braudaway and M. Yeung, "Effective and Ineffective Digital Watermarks", Proc. of IEEE International Conference on Image Processing ICIP-97,

Vol.3, pp. 9-12, 1997

[4] Mitchell D. Swanson, Bin Zhu and Ahmed H. Tewfik,“Multiresolution Scene-Based Video Watermarking using perceptual Models”, IEEE Journal on Selected Areas in Communications, Vol. 16, No. 4, May 1998.

[5] Watson, A.B. (1986). Temporal sensitivity. In Boff, K.,Kaufmann, L. & Thomas, J. (eds.), Handbook of Human Perception and Performance, 1, Chapter 6. New York: John Wiley and Sons.

[6] Andrew B. Watson, James Hu, John F McGowan III, “DVQ: A Digital Video Quality Metric based on Human Vison”, Journal of Electronic Imaging, 10(1), 20-29.

[7] M. Holliman, W. Macy, and M. Yeung, “Robust frame-dependent video watermarking,” Proceedings of the SPIE Security and Watermarking of Multimedia Contents, vol. 3971, pp. 186-197, San Jose, January 24-26, 2000.

[8] Zhihai He; Yong Kwan Kim; Mitra, S.K., "Low-delay rate control for DCT video coding via rho-domain source modeling," Circuits and Systems for Video Technology, IEEE Transactions on , Volume: 11 Issue: 8 ,Aug. 2001 Page(s): 928 –940

[9] R Lienhart, “Comparison of automatic shot boundary detection algorithms”. In SPIE Conf. on Storage and Retrieval for Image & Video Databases VII, volume 3656, pages 290–301, 1999

[10] Jianhao Meng and Shih-Fu Chang, "Embedding Visible Video Watermarks in the Compressed Domain," Proc. of ICIP International Conference on linage

Processing, Oct. 1998

Attached Files:

• dip01309.doc (280 KB)