伤城文章网 >  > Data Hiding in Film Grain

Data Hiding in Film Grain


Best Paper Award at the 2006 International Workshop on Digital Watermarking

Data Hiding in Film Grain
Dekun Zou, Jun Tian, Jeffrey Bloom, and Jiefu Zhai
Thomson Corporate Research 2 Independence Way, Princeton, NJ 08540, USA {dekun.zou, jun.tian, jeffrey.bloom, jiefu.zhai1}@thomson.net

Abstract. This paper presents a data hiding technique based on a new compression enhancement called Film Grain Technology. Film grain is a midfrequency noise-like pattern naturally appearing in imagery captured on film. The Film Grain Technology is a method for modeling and removing the film grain, thus enhancing the compression efficiency, and then using the model parameters to create synthetic film grain at the decoder. We propose slight modifications to the decoder that enable the synthetic film grain to represent metadata available at the decoder. We examine a number of implementation approaches and report results of fidelity and robustness experiments. Keywords: digital watermarking, data hiding, video watermarking, film grain, film grain technology.

1 Introduction
When a flat field is captured by an optical camera and transferred to film and that film is subsequently developed and printed, the result is not a flat field. The resulting field has a "grain" texture known as film grain. The grain pattern in a single frame is well modeled by band-limited Gaussian noise whose variance and pass-band are determined by the kind of film stock used, the development and printing processes applied and the brightness of the underlying imagery. Film grain is well known to Directors of Photography (DP) on a motion picture project. The DP selects the film stock and specifies the processing to obtain a desired film grain effect. The inspiration behind this paper is that the film grain pattern in a motion picture has a striking similarity to the white noise reference patterns typically used in digital watermarking. We investigate some possible techniques for replacing the natural film grain inherent in a motion picture with a synthetic film grain designed to encode some metadata about the content. While film grain enhances the feel of a motion picture, it also makes compression of that content more difficult as the pattern is noise-like and independent from that of the adjacent frames. The two obvious options are to spend more bits to maintain the film grain or sacrifice the film grain allowing it to be lost during quantization. However, recent work has provided a third alternative [2]. Special filters have been developed to estimate and model the film grain inherent in each frame. Once an estimate of the film grain has been obtained, it is removed leaving a version of the content that does not
Y.Q. Shi and B. Jeon (Eds.): IWDW 2006, LNCS 4283, pp. 197 – 211, 2006. ? Springer-Verlag Berlin Heidelberg 2006

198

D. Zou et al.

contain any film grain. This version of the content is easier to compress. The model parameters are then sent along with the compressed content to the decoder. At the decoder, a film grain synthesizer uses the model parameters to synthesize a film grain that has similar properties as the original, removed grain. The synthesized grain is then added back to the reconstructed imagery. This technique, known as Film Grain Technology (FGT), has been incorporated in a number of international standards and provides a platform for implementing a data hiding scheme. We begin this paper with an introduction to Film Grain Technology and then provide specific implementation details for a modification that can use the film grain to represent metadata. Finally, we present some experimental results and discuss next steps.

2 Film Grain Technology
Motion pictures are formed by exposure and development of photographic emulsion. They typically contain some kind of noise, which is called film grain. Film grain originates in the physical process of exposure and development of photographic film and exhibits quasi-random characteristics. Film grain forms in different intensities, sizes, and colors depending on the film stock, the developing and printing processes and the brightness of the underlying imagery. Human eyes can not distinguish a single grain; instead groups of these random patterns can be identified more easily. It is clearly noticeable at high resolution and therefore has become a distinguished trait that should be preserved during compression. Figure 1 gives examples of various film grain patterns that differ in intensity and grain size.

Fig. 1. Illustration of various film grain patterns

As we have mentioned, film grain is generally regarded as quasi-random pattern and contains very high entropy, which makes it very hard to compress. There are a few mathematical models which could be used to generate artificial film grain. In recent work, Gomila [2] presents Film Grain Technology as a new tool that allows

Data Hiding in Film Grain

199

encoding of film grain in motion pictures by means of a parameterized model which is transmitted as supplemental information. To support FGT, the MPEG-4 AVC standard has defined a film grain characteristics Supplemental Enhancement Information (SEI) message in the Fidelity Range Extension (FRExt) Amendment [3][4]. The FGT has also been adopted in HD-DVD as a mandatory part of the format specifications [5]. The basic framework of FGT is illustrated in Figure 2. At the server side, the input video is sent to film grain modeler and film grain remover1 simultaneously. The result of the film grain removal process is then compressed and the results of film grain modeler are encapsulated in SEI messages, which can be sent independently or embedded into compressed bit-stream for transmission. At the receiver side, if the SMPTE Specifications [4] are enforced, both the SEI messages and the output of the video decoder are sent to a film grain simulator, where bit-accurate film grain simulation is done. Bit accurate here has the same meaning as in video decoding specifications such as H.264/AVC: regardless of the means by which film grain simulation is implemented, the result of generated film grain should be exactly the same as the one generated by a reference implementation. A method for simulating film grain is preformed by generating random noise pattern and passing that pattern through a separable, 2D band-pass filter. Thus, the characteristics of a pattern are specified by the Gaussian variance (controlling the intensity of the grain) and the four cutoff frequencies of the band-pass filter which characterize the size of grain.

Input imagery Film grain removal / attenuation

Film grain modeler

SEI message

……. ……

Film grain simulator

To display

+
Video Encoder Video Decoder Compressed bit-stream

Fig. 2. Framework of Film Grain Technology (FGT)

If the SMPTE Specifications are enforced, in the context of the H.264/AVC, the synthesized film grain is generated and added on an 8×8 block basis. Thus, this frequency filtering model is performed block by block. Conceptually the generation of film grain is done in the following way: for each 8×8 block an 8×8 Gaussian random field is generated. The variance of the Gaussian random variable is controlled by the film grain intensity parameter, which characterizes the strength of the grain. The size of the grain is controlled by a frequency filtering using the four cutoff frequencies: lower and upper horizontal cutoff and lower and upper vertical cutoff. In practice, this filtering is often done in the DCT domain and the two band-pass filters are actually low-pass filters. Using low-pass instead of band-pass is a simplification in order to reduce complexity. Such simplification supports most of the applications
1

Film grain removal is actually an optional function as we expect the compression to essentially remove any film grain.

200

D. Zou et al.

without adversely affecting the visual quality. The intensity of the film grain pattern is actually dependent on the brightness of the underlying image data. The pattern intensity typically peaks at mid-intensity and decays in dark and bright regions. In very dark or very light regions, the pattern intensity can often be zero indicating that film grain is not visible in such extremes. One artifact of block-based film grain generation is that the full frame pattern can have blocking artifacts. To reduce these artifacts, a simple deblocking scheme is applied to the generated film grain. In the FGT specification for HD-DVD [4], the above steps are simplified by employing lookup table for low complexity implementation.

Fig. 3. Test frame: 704x480 crop from a frame of the StEM clip

(a) Original

(b) Compressed at 8 Mbps

(c) Film grain simulation

Fig. 4. Film grain simulation results on a close-up of one small part of the tested frame of Figure3

Figure 3 shows a 704×480 crop from a frame of the DCI-ASC mini-movie known as StEM (the original frame from which this is cropped is 1920×1080). In order to illustrate the film grain, we have further cropped a small region from the upper left and shown scaled up versions in Figure 4. Figure 4(a) shows the portion of the original picture. Figure 4(b) shows this same region after H.264 compression at 8 Mbps. As can be seen, the film grain pattern has largely been removed by the compression. The initial film grain has been modeled and simulated. Figure 4(c) shows the result after the film grain simulation is added to the image of Figure 4(b). Comparison of the three images in Figure 4 illustrates that the FGT approach can restore some of the film grain warmth to the compressed imagery.

Data Hiding in Film Grain

201

3 FGT-Based Watermarking Techniques
A simple approach to using film grain for data hiding is shown in Figure 5. The film grain in the input content is modeled and removed. The model parameters are used to generate one or many synthetic film grain patterns and the watermark payload is used to select or modulate these patterns before they are added back to the "cleaned" version of the content.
Payload Model Parameters

Original Content

Film grain Film grain modeler modeler

Film grain Film grain simulator simulator

Pattern selection Pattern selection and/or and/or modulation modulation

Film grain Film grain removal / / removal attenuation attenuation

“Clean” Content with film grain removed

+

Marked Content

Fig. 5. Simplified approach to film grain data hiding

This basic idea must now be placed in the context of the FGT framework that was illustrated in Figure 2. The main difference is that the "clean" version of the content is subject to a compression/decompression cycle before the synthetic film grain is introduced. In addition, the generation of the synthetic film grain, and thus the incorporation of a watermark payload, takes place at the client, not at the server. This is shown in Figure 6 below.
Payload

Original Content

Server

Film grain Film grain modeler modeler

Model Parameters

SEI message

Client
Film grain Film grain simulator simulator Pattern selection Pattern selection and/or and/or modulation modulation G

Film grain Film grain removal / / removal attenuation attenuation

“Clean” Content

Video Encoder Video Encoder

Compressed bit-stream

Video Decoder Video Decoder

I

+

I’

Marked Content

Fig. 6. Data hiding in the context of the FGT framework

Introduction of the watermark payload at the client has a number of implications on the applications for which this technology would be appropriate. First, there is a security implication as the client necessarily has access to an important part of the watermark embedder. Second, this approach now lends itself to applications in which the data to be embedded is available only at the client. For example, the client may introduce its own ID into the content. The client may introduce a time stamp indicating the playout time. The client may be redistributing the uncompressed content to a number of other processes and thus use the payload to identify each of those processes. The basic idea is to make the film grain pattern payload dependent. Several approaches are proposed. For this discussion, let I denotes a decompressed picture before film grain addition and let G denote a film grain pattern, which has the same

202

D. Zou et al.

dimensions as the original image. Finally, let I' denotes the image with film grain added. As discussed in Chapter 4 of Cox, Miller, Bloom [1], there are a number of ways to select or modulate patterns in order to represent messages. The direct message coding method designates a different pattern for each message. If the message, in this case, is the ID of the client, this can be accomplished by allowing each client to seed the film grain pattern generator with a different initial seed. This approach is examined further in Section 0. It is limited, however, to applications with few messages. For more flexibility in coding arbitrary payloads (up to a maximum length), an alphabet of symbols or film grain patterns is defined and multiplexing is used to combine the symbols. For illustration purposes, Section 0 presents a time division multiplexing approach where one symbol is embedded into each frame of the image sequence. Both approaches rely on the fact that the visual properties of the synthesized film grain pattern are based on the model parameters and do not depend on the specific key used to generate that pattern. The use of different seeds will result in different film grain patterns all with the same visual properties. Thus, data is embedded by selection of one or another visually equivalent pattern. 3.1 Direct Coding Using Unique Client Seed In FGT, the film grain patterns are generated block by block. For each block, a seed is used to generate the pattern. For one block, the seed is determined by the seed used in the previous block. A function is used to transform one seed into the next. The details can be found in the FGT specifications for various standards (e.g., [4]). The seeds are reinitialized after each SEI message is encountered. Therefore, once the initial seed is set, the whole pattern will be determined. Different clients can be assigned different initial seeds. The embedded film grain pattern will then be different for each user. The embedding procedure is illustrated in Figure 7.

Fig. 7. Direct coding FG embedding

For each frame, the detection process begins with a registration. For the current implementation we register with the original content, but other techniques can be used (see Section 8.3 of [1]). The original picture is then subtracted from the suspect picture. Here, the image used as the original is actually the decompressed clean image; Image I from Figure 7. The remaining difference is an estimate of the embedded film grain pattern. Since the film grain pattern generation process is deterministic given a particular initial seed and the clean content, and assuming that the initial seeds of each of the clients is known to the detector, a library of reference

Data Hiding in Film Grain

203

patterns can be generated at the detector. The reference pattern with the highest correlation to the extracted film grain estimate is the pattern most likely to have been embedded. This, in turn, identifies the most likely client. Figure 8 shows the detection procedure.

Fig. 8. Direct decoding of FG patterns

3.2 Space / Time Division Multiplexing Since the synthetic film grain patterns are generated on a block by block basis, there is the opportunity to embed a different symbol into each block. Using a symbol alphabet of size 2, this technique would embed one coded bit in each 8×8 block of image data corresponding to 5280 coded bits in each 704×480 frame or 32400 coded bits in a 1920×1080 frame. The data rate can be increased by a factor of n by defining a set of 2n different film grain patterns for each block. It is possible that the intensity of the film grain pattern in some blocks will be zero. This will be the case for very bright and very dark regions. Any information scheduled to be stored in these blocks would be lost. There are many ways to address this including reliance on error correction coding or protocols for skipping blocks that cannot reliable hold information. Alternatively, an entire frame or even groups of frames can be used to embed each symbol. Using groups of frames would provide some robustness to dropping of frames as can occur in advanced coding schemes. In the experiments of Section 0, we embed one symbol in each frame. In the remainder of this section, we present two different methods for embedding one bit per frame. Both of these methods are constructed around the specific details of FGT. Specifically, it has been mentioned that the synthetic film grain patterns are generated on a block by block basis and that the patterns are dependent on both the local intensity of the imagery and the cutoff frequencies specified in the model. In practice, these patterns are precalculated and stored in a database. For each combination of cutoff frequencies, the database contains an 8x8 array of 8x8 patterns. Allowing for a 4-pixel overlap in the horizontal direction, this yields 120 different patterns. Also available are the inverses of these patterns. The client uses a deterministic procedure to select one of the 240 available patterns for the block and then modulates that according to the local image intensity. The method that we present partitions the set of 240 patterns into two sets. One set represents a '0' value bit and the other set represents a '1' value bit. Two variations are presented. These two differ primarily in the assumption we make about information shared between the embedder and detector.

204

D. Zou et al.

3.2.1 Shared Seed The first variation presented assumes that the pattern selection function uses a seed value that is known by the detector. This would be the case when all clients use the same seed. The set of 120 8×8 film grain patterns are collectively referred to as the '1' set and the set of 120 inverse patterns are the '0' set. The seed value for each block is used to select one pattern from the '1' set. This continues for each block in the frame generating a full frame film grain image. If the bit for the current frame is also a '1', then the generated pattern is used. If the bit is a '0', then the inverse pattern is used. Since the detector has the same seed, it can also create the '1' pattern which is used as a reference pattern for correlation analysis. If the correlation is positive, then we conclude that a '1' bit was embedded. If the correlation is negative, then we conclude that a '0' bit was embedded. For a given block, the embedding is depicted in Figure 9 and can be approximated as in Equations (1), where Gk is the kth 8×8 pattern in the '1' set. This is only an approximation because a deblocking filter is subsequently applied across the blocks.

′ I k = I k + αGk where α = ? ?

?1 b = 0 ?+ 1 b = 1

(1)

α =?

?? 1 ?+ 1

b=0 b =1

Fig. 9. Multiplex FG embedding with a shared seed

The detection procedure is much simple than previous methods. Since the pattern we have added or subtracted is already known, the sign of the correlation between this pattern and the difference picture will determine the bit embedded in this block. Figure 10 illustrates the detection process. Each frame of the test sequence is registered and compared to the corresponding frame of sequence I. The difference is an estimate of the added film grain pattern and is correlated with the reference pattern for that frame. A positive correlation suggests a '1' bit and a negative correlation suggests a '0' bit.

? z <0→b =0 ? z > 0 → b =1

Fig. 10. Multiplex FG detection with a shared seed

Data Hiding in Film Grain

205

3.2.2 Detection with Unknown Seed There may be applications in which the detector cannot reproduce the pattern selection process used during embedding. This would be the case when each client uses a differ-rent seed for the selection process. To address this problem we partition the set of 240 8×8 patterns differently (recall there are 120 patterns in the database plus the inverse of each. Before introducing the partition strategy, let us first assume that the set has been appropriately partitioned into two sets of 120 patterns each, one set labeled the '0' set and one labeled the '1' set. To embed a '0', a film grain pattern from the '0' set is selected for each block of the frame. To embed a '1', a pattern from the '1' set is selected for each block of the frame. Each block has a different pattern, but all the patterns chosen for a particular frame come from the same set. In order to recover the bit without being able to reconstruct the selection process, the detector will correlate the film grain estimate from each block with each of the 240 reference patterns in the database2. The pattern that yields the highest correlation is the pattern most likely to have been embedded. This process is repeated for each block and all of the results are combined (e.g., by voting) to obtain the recovered bit value. If we were designing the database of patterns for this purpose, we would want each of the patterns in the '0' set to be orthogonal to all of the patterns in the '1' set and each of the patterns in the '1' set to be orthogonal to all of the patterns in the '0' set. This would minimize the likelihood that a block in which a '1' had been embedded ends up having a high correlation with a '0' pattern and vice versa. However, we do not have the luxury of designing the database. It has already been designed and is written in the FGT specification. The next best thing is to partition the set of 240 patterns so as to minimize the maximum correlation across the two sets. Two patterns with high cross correlation should be placed in the same set. For block k, the embedding can be approximated by Equation (2) where Gk0 denotes a randomly selected pattern from the '0' set and Gk1 denotes a pattern randomly selected from the '1' set. Again, this is an approximation because a deblocking filter is subsequently applied across blocks. Figure 11 illustrates the embedding process.
? I + Gk 0 ′ Ik = ? k 1 ? I k + Gk b=0 b =1

(2)

Fig. 11. Multiplex FG embedding with a unique seed

The detection procedure is illustrated in Figure 12. The original movie is used for both geometric/temporal registration and subtraction. Each extracted difference image
2

In fact, only 120 correlations are necessary since the correlation with one pattern will have the same magnitude and opposite sign as the correlation with the inverse of the pattern.

206

D. Zou et al.

is divided into blocks and correlation detection will determine which pattern was most likely embedded.

Fig. 12. Multiplex FG watermark detection without knowledge of embedding seed

4 Properties
The approach presented here results in watermarked imagery that has the same fidelity as FGT itself. The FGT film grain estimation, modeling, and synthesis process results in imagery with convincing film grain after decompression. Each frame, and even each block within each frame, contains a different, signal independent film grain pattern. The data hiding approach simply takes control over the synthesis process, forcing the output to specific states specified by the data payload rather than allowing it to be completely random. In Section 0 we describe the experimental method used to confirm the fidelity. Note that the noise-like watermark pattern is not designed to be imperceptible. It is designed to be indistinguishable from the synthetic film grain pattern normally introduced by FGT during decompression. The detection presented here is an informed detection. A version of the original content, with film grain already removed, is used for registration. In addition, it is used to remove as much of the host image as possible leaving an estimate of the added synthetic film grain for detection. This detection process is described more fully in Section 0. Note that generation of the reference patterns requires knowledge of the film grain model parameters obtained from the original content as well as intensity information from the original content and, potentially a key unique to the particular title. Therefore, even with an alternative registration approach, this method does not lend itself to blind detection. Film grain watermarking has proven to be robust to a number of different processes. In Section 0, we present data for embedding efficiency as well as robustness data for distortions due to compression and noise removal that both tend to remove film grain. Since the reference patterns are published in the FGT specification, it will be difficult to create a secure application using the multiplexing methods of Section 0. However, the direct coding method has more potential for security. In general, we recommend use of the methods presented here for applications that do not require security against unauthorized removal.

Data Hiding in Film Grain

207

5 Experiments
The shared seed temporal multiplex approach of Section 0 was implemented because it is the closest to the existing FGT standard requiring the fewest changes to the existing film grain generators. Parameters are estimated and the resulting imagery is examined by viewing experts. If any block on any frame has an undesirable synthetic film grain appearance, the parameters can be adjusted. The result of this tedious process is a predefined set of patterns to be used for each frame. Every client process will produce the exact same, "bit accurate", reconstruction as that approved by the viewing experts. The premise used in these experiments is that use of the inverse pattern will be equally acceptable. Informal fidelity assessments by our local film grain experts support this premise. Both fidelity and robustness of proposed FGT based video watermark methods were tested using 1437 frames from the DCI-ASC Mini-movie, StEM. The resolution is 704×480. 5.1 Fidelity For fidelity assessment we compare a sequence with standard FGT film grain (not representing any payload data) with a watermarked sequence in which the film grain does represent payload data. Clearly, this experiment need not be done but only to verify that there are no unanticipated consequences. Viewers included 2 film grain experts and 4 other viewers experienced in viewing compression artifacts, but not necessarily film grain artifacts. Side-by-side, the differences in film grain could not be perceived and none of the testers could identify which was the watermarked sequence and which was the sequence with standard film grain. For illustration purpose, consider the image in Figure 4(a) showing an enlargement of a part of the original image with its natural film grain. The same region from a watermarked version of the frame is shown in Figure 13(a). Compare this with the picture in Figure 4(c), where a standard film grain pattern is added.

(a) Marked frame

(b) 640 Kbps compressed

(c) after noise removal

Fig. 13. Frames after watermark embedding

5.2 Embedding Efficiency In order to assess the efficiency of the embedder, the embedder output is fed directly into the detection process without any additional distortions. The payload for this

208

D. Zou et al.

experiment was a random bitstream with an equal number of 1's and 0's. The detection measure used is normalized correlation which has a range of -1 to +1. The reference pattern used is a full frame constructed from the preselected reference blocks corresponding to a 1 bit. A histogram of the resulting correlation values for 1437 frames is shown in Figure 14.

Fig. 14. Watermark efficiency using a reference pattern constructed from the preselected reference blocks corresponding to a 1 bit

Fig. 15. Watermark efficiency using a deblocked positive reference pattern for detection

Since an equal number of 1s and 0s are embedded, two sharp peaks can be observed in the histogram. The magnitudes of the correlation values are slightly smaller than 1. This is unlikely due to clipping as very dark and very light regions do not get film grain added. We suspect that the reason is that the reference pattern we used is the film grain pattern before deblocking. Therefore, it is slightly different from the pattern actually added. Note that the deblocking filter applied to the inverse pattern does not result in the inverse of the result of the deblocking filter applied to the positive pattern. Therefore, we would need to generate two reference patterns to correctly address this problem. To test this hypothesis, the deblocking filter is applied to the positive reference pattern and the correlations are again performed. The results of this test are shown in Figure 15. As expected, the magnitudes for the 1's case have moved up to +1. Use of the deblocked inverse reference pattern will have the same effect on the 0's case. 5.3 False Positive Probability In this presentation, we do not provide a formal false positive analysis. We do, however provide a preliminary look at the detection values obtained when the content has not been watermarked. The test material is the original movie clip with its natural film grain. This is provided as input to the detector. Again the number of frames in this test is 1437. The distribution of correlation values is shown in Figure 16. In this small test, no frame yielded a detection value with magnitude greater than 0.01. If these values are typical for various reference patterns and content, we can expect to be able to set a correlation threshold that safely distinguishes between marked, undistorted content (Figure 14) and unmarked content (Figure 16). To confirm this, a larger false positive analysis must be performed.

Data Hiding in Film Grain

209

Fig. 16. Correlation distribution with unmarked content. Note the x-axis scale. No frame yielded a detection value with magnitude greater than 0.01.

5.4 Robustness In this presentation, we examine the robustness of the watermark to lossy compression and Gaussian noise removal. An H.264 encoder/decoder was used for lossy compression. Various compression bitrates were tested from a maximum of

(a) 6692 Kbps

(b) 1640 Kbps

(c) 880 Kbps

(d) 535 Kbps

Fig. 17. Distribution of detection values after compression with various bit rates

210

D. Zou et al.

11 Mbps to a minimum of 470 Kbps. Examples of the visual impact of these distortions are shown in Figure 13(b) for compression and Figure 13(c) for noise removal. Perceptually, the film grain has been greatly reduced. The distribution of detection values for some of the compression cases are shown in Figure 17. These figures show that the magnitudes of the detection values decrease with increased compression. Note that the x-axes are different for the different figures. The average detection magnitudes in these tests are summarized in Table 1. At 535 Kbps, the correlation values begin to overlap the levels seen for unmarked content. This suggests that we may not be able to distinguish, with certainty, between an unmarked sequence and a sequence that has been marked and significantly compressed. At least we may not be able to make this distinction based solely on the detection values. Advanced message coding and the use of error correcting codes can improve the certainty. However, despite these low detection values, it is important to note that all of the correlations have the correct sign and thus each of the 1437 bits was correctly recovered in each of the robustness experiments performed. The Gaussian noise removal did remove the visual noise, but enough of the film grain pattern remains for solid detection. The distribution of detection values for the 1437 frames is shown in Fig. 18. Again, all bits were correctly recovered.
Table 1. The mean magnitude of the detections value decreases along with compression bitrate

Compression Bit Rate (Kbps) 11741 6692 1640 880 535 477

Average Detection Value 0.7493 0.5841 0.1256 0.0458 0.0207 0.0167

Fig. 18. Distribution of detection values after Gaussian noise removal

Data Hiding in Film Grain

211

6 Conclusion and Next Steps
In this paper, we have presented a data hiding technique custom designed for applications using Film Grain Technology. Although FGT may not be widely known to the watermarking community, it has been incorporated in a number of important international compression standards and will become a widely distributed technology. This work extends the utility of FGT for data hiding applications. The fidelity of the approach is established by the inability of expert viewers to correctly classify watermarked sequences and standard film grain sequences. This fidelity analysis was performed on a small sample of imagery and must be expanded. The work presented here took a preliminary look at the false positive characteristics. This was helpful in providing a context for assessing the robustness results. However, a more thorough false positive analysis is required. Such an analysis will examine much more content of differing types and a number of different reference patterns. The resulting histograms can be modeled and the models can be used to establish the relationship between the threshold and the probability of a false positive. In addition to such an empirical study, an analytical study that characterizes the distribution of detection values is desired. In the robustness experiments described, all of the bits were correctly recovered despite low detection values. Additional robustness experiments are needed to identify the limits at which the bit error rate becomes non-zero and examine the interaction of those errors with an appropriate error correcting code. In addition, we need to assess the robustness to other types of distortions.

References
[1] I. Cox, M. Miller, and J. Bloom: Digital Watermarking: Principles & Practice, San Mateo, CA: Morgan Kaufman, 2001. [2] C. Gomila: SEI message for film grain encoding. Contribution JVT-H022 to 8th JVT Meeting, Geneva, Switzerland, 23-27 May, 2003. [3] ITU-T Recommendation H.264 | ISO/IEC 14496-10 International Standard with Amendment 1. [4] SMPTE RDD 5-2006: Film Grain Technology - Specifications for H.264 | MPEG-4 AVC Bitstreams. [5] HD DVD-Video Specifications for High Density Read-Only Disc, Version 1.0, August, 2005.


搜索更多“Data Hiding in Film Grain”

网站地图

All rights reserved Powered by 伤城文章网 5xts.com

copyright ©right 2010-2021。
伤城文章网内容来自网络,如有侵犯请联系客服。zhit325@126.com