CROSS-REFERENCE TO RELATED APPLICATIONS

This is a divisional application of U.S. application Ser. No. 08/829,524 filed Mar. 28, 1997.

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

Not applicable

BACKGROUND OF THE INVENTION

The present invention relates to data transmission, and more particularly to a method and apparatus for transparently embedding data within a video signal in order to indicate authentication, program ownership and/or reception-verification of the video signal.

Television signals are usually copyrighted or otherwise proprietary to the originator and, in the case of television network distribution, are distributed to affiliate stations for re-broadcast. Unauthorized re-broadcast is difficult to detect since it is often difficult to determine the creator or originator of the material from the material itself. This is particularly true for short sequences or when the video image has been cropped or modified to mask any identification logos that may have been explicitly burned into the video image.

Another common problem with television and video distribution, which is becoming more severe, is the synchronization of an aural or audio component that is to be distributed with the video signal. Modem digital signal processing techniques requiring large buffering, such as MPEG compression, add latency to the video distribution and, since the audio may be distributed or processed separately, an error in audio to video “lip-sync” or in sound to action often occurs. Sometimes this latency is variable, requiring continual re-synchronization. In addition to audio to video synchronization problems, the audio signal may be distorted or missing entirely. By embedding, for example, the audio envelope as data within the video signal, it is also possible to detect and compare the received audio with the original which was coded and embedded as data within the video signal as a quality measure as well as video to audio delay.

Another problem is with authentication of video signals representing visual images, possibly created by a computer, where it is desirable to detect a forgery or imposter signal. In this case the video signal may be replicated to such a degree that it is difficult to detect it from the authentic video sequence. Also it is sometimes desirable to detect the beginning and/or ending of a particular sub-sequence of a video signal, such as a motion test sequence for in-service video quality assessment of video processed by MPEG compression. It is possible to devise a remote receiver to capture a special test sequence segment of a distributed video signal and compare that segment to an undistorted, stored version to assess the quality of the video. Although it is possible to uniquely and reliably detect when the sequence occurs so that it, and only it, is captured, this generally requires the advance processing of a signature vector, transmission of that vector through a separate channel and a correlation means at the receiver. By transparently embedding data into the video a predetermined start and end code may be detected by the receiver without any preloading of signature vectors and signature vector preprocessing.

Published Canadian Patent Application No. 2,174,413 (A1) by Geoffrey B. Rhoads entitled “Identification/Authentication Coding Method and Apparatus” discusses techniques for providing authentication of image signals. In Rhoads an imperceptible N-bit identification code is embedded throughout an image with a small noise pattern in a coded fashion. In particular bits of a binary identification code are referenced sequentially to add up to N independent, noise-like patterns to the original image signal. The detection of these patterns is done by N sequential correlations with stored replicas of each pattern. This may also be done simultaneously by N correlators, as is well described in the public domain as a “correlation receiver.” Rhoads further discloses adding or subtracting exactly N independent, noise-like images to improve detection and/or encoded image quality. This later modification, referred to by Rhoads as “true polarity”, is well described in the public domain as “bi-orthogonal signaling” for a correlation receiver allowing 2{circumflex over ( )}N composite symbols (patterns) to be created by adding or subtracting N bi-orthogonal symbols (patterns). A disadvantage of both of Roads' methods is that each of the N noise-like patterns, which are added or subtracted to form one of the 2{circumflex over ( )}N possible composite patterns, needs to be properly scaled and designed to minimize image degradation by the addition of the composite pattern to the image. A further disadvantage is that the original unencoded image, as well as the N patterns, need to be stored in the receiver so that the unencoded image may be subtracted from the encoded image for detection.

U.S. Pat. No. 4,969,041 issued Nov. 6, 1990 to William J. O'Grady and Robert J. Dubner entitled “Embedment of Data in a Video Signal” discloses adding one of a plurality of low-level waveforms to the video signal, with the level of the low level waveform being below the noise level of the video signal, for authentication purposes or for transmitting information. Each low-level waveform corresponds to a particular data word being embedded. At the receiving end the video signal is correlated with an identical set of low-level waveforms to produce a set of correlation coefficients—the highest correlation coefficient indicating the presence of a particular one of the low-level waveforms which is then converted into the corresponding data word. The Rhoads composite pattern consists of the summation of up to N independent patterns, which would be the same as O'Grady/Dubner when N=1 since only one pattern is sent at a time. For an N-bit data word O'Grady/Dubner implies the need for storage of 2{circumflex over ( )}N patterns rather than the N patterns of Rhoads. But since 2{circumflex over ( )}N patterns may be generated by summing N bi-orthogonal sub-patterns, the effect is the same and only N patterns need to be stored in a manner identical to Rhoads. In both of these patents the symbol rate is one pattern per picture or field.

Some other drawbacks of O'Grady/Dubner are that it limits the degree to which a pattern or sequence of patterns can be hidden by not fully exploiting the limitations of the human psychovisual process, such as spatial masking; it does not fully exploit the available signal to noise ratio since the correlation output is unipolar; and it does not include coding of the data represented by the embedded symbols so as to provide temporal redundancy for error correction over missing video frames.

What is desired is a method of embedding an unobtrusive data pattern into the video signal that is useful in addressing the above-noted problems.

BRIEF SUMMARY OF THE INVENTION

Accordingly the present invention provides a method and apparatus for transparently embedding data into a video signal by modulating a particular carrier frequency with a randomized pattern representing the data. The randomized pattern may be a combination of the data with the output from a pseudo-random binary sequence generator, or the data may be converted first into a bi-orthogonal pattern selected from among a plurality of such patterns and then randomized. The resulting modulated carrier frequency is then amplitude modulated based upon a human visual model so that the modulated carrier frequency amplitude is greater in areas of high image complexity and lower in areas of low image complexity. The subliminal data represented by the modulated carrier frequency is added to the video signal. At a decoder the data is recovered using appropriate correlation techniques by first filtering the video signal to increase the subliminal data component relative to the image component, then demodulating the data component to produce the data pattern, and then correlating the data pattern with the possible data patterns to identify the particular pattern and associated data word.

The objects, advantages and other novel features of the present invention are apparent from the following detailed description when read in conjunction with the appended claims and attached drawing.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING

FIG. 1

is a block diagram view of an encoder for transparently embedding data in a video signal according to the present invention.

FIG. 2

is a block diagram view of a feedback model for determining parameter values for a human visual model according to the present invention.

FIG. 3

is a block diagram view of a decoder for extracting embedded data in the video signal according to the present invention.

DETAILED DESCRIPTION OF THE INVENTION

The invisibility of embedded data in a video signal is achieved by several independent factors acting together. The first factor is the principle of spectral spreading from Direct-Sequence Spread Spectrum (DSSS) modulation of a carrier frequency. The low-bandwidth/low-rate data patterns or symbols are modulated by a higher rate Pseudo Random Binary Sequence (PRBS), spreading the power spectrum around the carrier frequency so that the time-domain waveform appears noise-like or lacks the spatial coherence that the eye/brain is good at seeing even at low levels. Also by careful selection of the carrier frequency and phase it is possible to further reduce the visibility of the resulting spatial pattern created by the PRBS modulated symbols (patterns) on a DSSS modulated carrier.

Secondly it is also desirable to filter the embedded signal both vertically and horizontally to reduce edge visibility, further enhancing it's overall invisibility. A third factor that is significant is to exploit visual masking by varying the data signal level according to the video content.

Another factor is the selection of a Field Alternate Carrier (FAC) frequency and phase which becomes the DSSS modulated carrier. The Invisibility of this carrier is a function of it's frequency, but it's phase from one field to the next is very important. A first-order cancellation occurs by repeating the entire data pattern on the even field (field

1

) again on the odd field (field

2

), but with an inverted carrier phase, causing an interlaced line pair cancellation. By repeating the data in each field, an even-line field comb filter may be used in a decoder to cancel the field to field correlated video and recover the data pattern with a 3 dB advantage at a cost of a 50% data-rate penalty. For progressive scan video systems the Field Alternate Carrier frequency and data repeat may be replaced by a Line Alternate Carrier frequency and data repeat. This gives the same or better visual cancellation and has the same data rate as the interlaced scan system.

With video signals the decoder is greatly simplified since carrier, code and data lock are simply derived from the abundant timing information in sync and burst. In addition to the rejection afforded by the comb filter, a bandpass filter is used to suppress the large carrier components of the program video signal. After the PRBS correlation, or de-spreading, various detection methods are possible. An integrate and dump matched filter driving a software, bi-orthogonal M-ary scheme may be used. Two options with and without channel coding also may be used. Numerous alternatives include combining convolution coding with Walsh code assignments (Ungerboech TCM) and concatenated RS and convolution.

Channel coding is best done in a holographic, two-dimensional (2D) sense by shuffling or the like for robustness against video motion of pixel segments passed through the comb filter. Such 2D holographic shuffling of data may be extended to three-dimensional (3D) shuffling because video has a time sense unlike still pictures. This may be done by repeating some of the data over several frames, which could be valuable for video sequence identification. The data that identifies the video sequence may be given a lower energy, i.e., fewer pixels, than the data that identifies the particular frame within the video sequence, since the frames come only once while the sequence has many frames in it. Data security through data encryption may always be applied separately from the channel coding, allowing the decoder design to be public or even standardized.

Referring now to

FIG. 1

an encoder

10

is shown having an analog video signal which is input to an anti-aliasing, lowpass filter

12

, the output of which is input to a sync/burst separator

14

to generate necessary timing signals and to an analog to digital converter (ADC)

16

to produce a digital video signal. A sample clock phase locked loop (PLL)

18

in response to one of the timing signals (subcarrier frequency fsc) from the sync/burst generator

14

provides a sample clock at four times subcarrier (4 fsc). Where the input signal is a digital video signal, the timing signals are extracted from the digital video signal and the lowpass filter

12

and ADC

16

are not needed. The digital video signal is input to an adder

20

where a subliminal data signal is added to the digital video signal. The resulting digital video signal is input to a digital to analog converter (DAC)

22

to produce a modulated analog video signal which in turn is processed by an output lowpass reconstruction filter

24

to produce a final encoded analog video signal. For digital distribution the DAC

22

and filter

24

may be eliminated.

A modulator circuit

30

generates the subliminal data signal for input to the adder

20

. The modulator circuit

30

has as inputs the vertical and horizontal timing signals from the sync/burst separator

14

and the desired subliminal data to be added to the digital video signal. The subliminal data is input to a circular shift register

32

which has as inputs the horizontal and vertical sync signals. The subliminal data from the circular shift register

32

may be input to an exclusive OR gate

35

where it is combined with the output from a pseudo random binary sequence (PRBS) generator

34

. The PRBS generator

34

is reset by the vertical sync signal and clocked by a clock signal derived via a local PLL

36

and a divider

38

. The output from the exclusive OR gate

35

is a series of +/−1's which is then input to a mixer

40

. A numerically controlled oscillator (NCO)

42

provides a modulator frequency fo that also is input to the mixer

40

. The output from the mixer

40

is a modulated frequency signal which is the DSSS signal as modulated by the output from the PRBS generator

34

and the data from the circular shift register

32

.

To provide the Field Alternate Carrier frequency an inversion of the phase of either of the inputs to the modulator

40

is performed using a divide-by-two circuit

37

having the vertical interval timing signal as an Input followed by another exclusive OR gate

39

. The exclusive OR gate

39

has as shown the other input coupled to the output of the first exclusive OR gate

35

, the output being coupled to the modulator

40

. Alternatively the exclusive OR gate

39

may be inserted between the NCO

42

and the modulator

40

, between the circular shift register

32

and the first exclusive OR gate

35

, or between the PRBS generator

34

and the first exclusive OR gate. For progressive scan systems the input to the divide-by-two circuit

37

would be the horizontal interval timing signal so that the Line Alternate Carrier frequency is implemented. In either case alternate lines in each complete frame of the video image sequence have the same data, but inverted in phase, so that subtraction of alternate lines by a comb filter in a decoder enhances the subliminal data signal.

The modulated frequency is input to two-dimensional, horizontal and vertical spatial frequency, softening filters

44

. The horizontal filter

44

A is centered about the modulator frequency fo to mitigate any fast edges in the modulated frequency signal, and the vertical filter

44

B mitigates fast vertical edges (line-to-line transitions). The modulated frequency signal is the subliminal data signal that is input to the adder

20

via a blanking gate

46

so that the addition occurs only during the active video portion of the digital video signal.

A human vision model (HVM) circuit

48

has as inputs the digital video signal at the input to the adder

20

. One practical implementation of visual masking based upon the human visual model is shown. A first aspect of the human visual model is noise tolerance in areas of high spatial detail, and this is controlled by a two-dimensional highpass filter

41

, such as a Sobel filter, followed by an envelope squarer

43

and a first control amplifier

45

. A second aspect is sensitivity to noise at different local contrasts, and this is controlled by a two-dimensional lowpass filter

47

followed by a transfer offset circuit

49

and a second control amplifier

51

. The parameters A

1

and A

2

represented by the first and second control amplifiers

45

,

51

are adjusted for a desired minimum data pattern visibility over a range of picture material. A third aspect may be that of temporal masking

55

where noise on screen cuts or motion is temporarily less visible. The resulting outputs from these paths, describing the complexity of the image rendered by the digital video signal, are summed by a summer

53

to produce a gain signal to control a multiplier

50

to which the modulated frequency signal is input prior to the adder

20

. Thus for portions of the digital video signal where the image is unchanging in motion and/or has low spatial detail and contrast or color/shading variations, the gain is relatively low; and where the image is complex due to motion and/or high spatial detail, the gain is relatively high. In fact the amplitude of the subliminal data signal does not need to be less than the noise floor in complex regions of the video digital signal to retain an acceptable level of invisibility.

A feedback model

60

for determining the parameter values A

1

, A

2

represented by the first and second control amplifiers

45

,

51

and the transfer curve of the transfer-offset circuit

49

is shown in

FIG. 2. A

desired or acceptable distortion value is pre-set and dynamically maintained by feedback based on the human visual model. An image sequence is stored in a video buffer

62

and repeatedly output in order to compare the output from the human visual model

64

prior to the encoder

10

with the image sequence derived from the output of the human visual model

66

following the data encoder. The comparison between the human visual model outputs is performed by means of a means absolute difference or mean square error algorithm

68

. The resulting difference or error is input to a comparator

69

to which also is input a distortion level which is adjustable to the desired level of invisibility. The parameters of the first and second control amplifiers

45

,

51

and the transfer curve of the transfer-offset circuit

49

are adjusted to maintain comparator

69

output high. This is repeated over a range of exemplary image sequences. This then produces an acceptable estimate of these parameters which may be fixed for all video to be encoded. The human visual model is only as elaborate as necessary to match the dominant elements of the human visual system as it relates to the desired level of masking.

This feedback model may also be used without the masking modulator

48

in an iterative manner to directly determine the pattern level gain signal for input to the multiplier

50

. In other words a vector or array of gain signal values, maximized for each pixel, may be determined based on the human visual model feedback, optimizing the root mean square (RMS) value of the embedded pseudo-noise pattern while maintaining the desired level of distortion.

Forward error correction may be added to the modulator

30

by inputting the subliminal data to a forward error correction (FEC) circuit

52

prior to input into the circular shift register

32

. This will provide error correction for bit or byte errors when the data from the circular shift register

32

is input directly to exclusive OR gate

35

. The FEC circuit

52

may use a bit-error correction code, such as Golay codes, as the forward error correction algorithm to improve the data integrity for bit errors, or a more sophisticated channel coding such as concantenated convolution and Reed-Solomon (RS) code may be used, as is well known in the art. These coding techniques combined with bit shuffling provide error correction across frames so that entire frames may be lost without any loss in data. For example FEC

52

may be designed to provide RS word-error correction where each frame contains a coded binary word of a fixed length of m data bits, allowing missing words (frames) to be completely recovered from the redundancy in adjacent words (frames). In this case the output from the circular shift register

32

may be input to an orthogonal pattern selector

54

, such as a Walsh encoder, to produce the subliminal data coded to select one of a set of orthogonal patterns input to the exclusive OR gate

35

. Walsh codes are one-dimensional orthogonal sequences used in the well-known IS-95 Code Division Multiple Access wireless telephone standard. Also Walsh sequences are the basis for a particularly well-known set of two-dimensional orthogonal patterns called Haadamard patterns. Orthogonal patterns are those that, when correlated with all other patterns in the set, produce a positive correlation coefficient for only one of the patterns, with all of the other patterns producing a zero correlation coefficient. Where each pattern represents m bits, a total of 2{circumflex over ( )}m patterns are required and the data from the circular shift register

32

acts as an address to the pattern selector

54

. By using bi-orthogonal patterns only 2{circumflex over ( )}(m−1) patterns or sequences are needed for m bits. These patterns may be input directly to the mixer

40

without the need of exclusive OR gate

35

, but the bi-orthogonal patterns or Walsh sequences may not all appear visually random, which would make them more visible when added to the program video signal. However when this pattern is exclusive OR'd with the PRBS generator output, the resulting light pattern at the output of the modulator

40

becomes much more random in appearance and is difficult to distinguish from noise. The resulting modulation is a randomized bi-phase modulation spreading the spectral energy around the carrier. Circular shift register

32

allows the data or data pattern to be repeated on interlaced fields so that the visual line pair cancellation and field comb data separator in the decoder may be used, as described above.

A decoder

70

is shown in FIG.

3

. The input section

72

is the same as for the encoder

10

, having a lowpass anti-aliasing filter

74

, a sync/burst separator

76

to generate the timing signals, a PLL

78

to generate the sample clock and an analog to digital converter

80

to produce the digital video signal. The decoder

70

also includes a Local Oscillator (LO) section

82

which includes a PRBS generator

84

with PLL

86

and divider

88

and a numerically controlled oscillator (NCO)

90

that produces the demodulation carrier frequency fo. The outputs from the PRBS generator

84

and the NCO

90

are input to a mixer

91

to produce a pseudo random demodulated LO signal (BPSK—Bi-Phase Shift Keyed) for a down-converter mixer

102

.

The digital video signal is input to preprocessing filters

92

that Include a bandpass filter

94

centered at the modulator frequency and a high pass field-comb filter

96

including a field delay

98

and a subtractor

100

. These filters remove much of the program video signal, enhancing the detection of the subliminal embedded data signal. The field-comb filter

96

is particularly effective since the embedded data pattern is repeated with inverse polarity on adjacent fields, doubling it's level at the filter output while the program video signal that is correlated on adjacent fields is removed. The filtered digital video signal is input to the down-converter mixer

102

and mixed together with the pseudo random modulated LO frequency signal. The resulting baseband digital signal is input to a correlation/matched filter circuit

104

.

The correlation/matched filter circuit

104

has as inputs the baseband digital signal and the horizontal and vertical interval timing signals. The timing signals are input to an orthogonal pattern generator

106

, such as a Walsh generator, that is comparable to that in the encoder

10

. The timing signals are also input to a matched filter

108

to which also is input the baseband digital signal. The matched filter

108

includes an integrate and dump circuit

110

followed by a sampler

112

. If the embedded data is only as bits encoded with an FEC channel code, as discussed previously, matched filter

108

output may be converted to binary data by slicer

124

and input directly to FEC decoder

126

to output the decoded and error corrected subliminal data. In the case where the data is coded as bi-orthogonal patterns, the output from the sampler

112

is input to a bi-orthogonal correlation receiver

114

that includes a correlation mixer

116

for each of the patterns produced by the pattern generator

106

so that the sampler output is correlated with each pattern. The outputs from the correlation mixers

116

are integrated by respective integrators

118

, the outputs of which are input to a maximum magnitude selector

120

. Due to the orthogonality of the patterns only one of the outputs from the integrators

118

has a significant magnitude that is detected by the magnitude and magnitude polarity selector

120

as one of the predetermined m bit patterns or data words assigned to each pattern. These m data words may have been FEC coded as described previously. In this case the error corrector

126

is designed to correct word errors based upon the redundancy in other decoded data words in order to correct data errors from missing video frames or from heavily distorted embedded data patterns. The magnitude selector

120

detects not only the greatest magnitude, but the polarity of that magnitude to provide an extra bit. In other words eight patterns represent three bits plus one bit for the eight inverted patterns, providing m=4 bits.

The subliminal data channel provided by the present invention may be used for stamping an identifying mark on a video sequence even if some of the video frames have been lost or portions of video frames have been cropped or distorted through lossy compression/decompression processing. The identifying mark may represent an audio signal or at least a signature of the audio into the active video to allow a remote video receiver to compare the audio signature with the received audio to automatically advance or delay the relative audio to video timing to maintain sound to video synchronization as encoded at the source. The subliminal data channel may represent modem cryptographic digital signature authentication in the video. It may be used to identify a motion test sequence, or portion thereof, to be captured for video quality assessment, such as discussed in co-pending U.S. patent application Ser. No. 08/605,241 filed Feb. 12, 1996 by Bozidar janko and David K. Fibush entitled “Programmable Instrument for Automatic Measurement of Compressed Video Quality”. The subliminal data channel also may be used to analyze timeliness of service, i.e., did the video get aired at the desired time in its entirety?

Thus the present invention provides for transparently encoding data into a video image by modulating a particular carrier frequency, providing interlaced field visual masking with a unique one of 2{circumflex over ( )}(m−1) bi-orthogonal patterns, the pattern representing m bits, randomizing the pattern with a PRBS sequence to reduce visibility, amplitude modulating the random pattern modulated carrier to further maximize embedded signal amplitude and minimize visibility based on a human visual model exploiting visual masking, and adding the resulting pattern modulated frequency signal to the video such that the data is recoverable using appropriate correlation techniques.