Patent Application
20130016911
The invention relates to the general field of protecting digital content from illegal copying and of identifying the source at the origin of an illegal copy. The domain of the invention is also known under the name of traitor tracing or user forensic.
More particularly, the invention relates to a method for classifying projection recaptures wherein a copy of a video content is classified as recorded from a digital projection of the video content or as recorded from an analog projection of a celluloid print of the video content.
This section is intended to introduce the reader to various aspects of art, which may be related to various aspects of the present invention that are described and/or claimed below. This discussion is believed to be helpful in providing the reader with background information to facilitate a better understanding of the various aspects of the present invention. Accordingly, it should be understood that these statements are to be read in this light, and not as admissions of prior art.
Among sources of illegal copy, the capture of a movie performed with a digital recording device (or camcorder) in a movie theater is largely widespread. To deter such in-theater piracy, forensic marks that survive camcording are embedded in projected movies so as to identify the specific cinema where illegal capture occurred. In digital cinema, digital watermarks carrying the theater identity as well as projection date and time are embedded at playback time. In legacy cinema projecting celluloid prints, dots arranged in a unique pattern so as to identify a particular print of a movie are printed on several frames during the manufacturing process.
Thus, when an illegal copy is analyzed, the forensic process to be applied depends on the type of leakage: camcorder capture in a digital cinema, camcorder capture in a conventional cinema, but also DVD rip and others. Classification of incoming illegal copies according to their origin, prior to further specific in-depth forensic analysis, is therefore needed to the whole forensic process. If the decoding of forensic marks is partly or fully automated, there is no such automated way of classifying incoming illegal video copies depending on their origin, especially for projection recaptures copies. Today, visual inspection and manual classification is used to characterize illegal camcorded video content during theatrical movie projections.
Thus known methods for forensics analysis raise the issue of distinguishing illegal copies recorded in digital cinemas and illegal copies recorded in legacy cinemas (celluloid film).
A method for automatically characterizing illegal camcorded video content during theatrical movie projections is therefore needed. Features automatically extracted from the illegal video characterizing whether the illegal video was recorded from a projection of a celluloid film or from a projection of a digital copy are further needed.
To this end, the invention relates to a method for classifying projection recaptures wherein a copy of a video content is classified as recorded from a digital projection of the video content or as recorded from an analog projection of the celluloid print of the video content. The method is remarkable in that the method is fully automated and that the step of classifying is a function of at least one feature among (i) spatial illumination uniformity, (ii) on-screen picture stability, (iii) temporal illumination pulse extracted from a digital representation of the copy. The invention is advantageously well adapted to automatically characterizing illegal video content camcorded during theatrical movie projections. The goal of the invention is to be able to distinguish between illegal copies recorded in digital cinemas and illegal copies recorded in legacy cinemas. The motivation for building such an oracle is to provide valuable insight at the beginning of the forensics analysis in order to switch between different forensic processes accordingly.
In a first preferred embodiment the method comprises a step of extracting the spatial illumination uniformity from the digital representation of the copy, the step of extracting the spatial illumination uniformity further comprising computing a representative frame of the video by temporally averaging luminance frames of the digital representation of the copy. According to particularly advantageous characteristics, the step of extracting the spatial uniformity further comprises cropping-out parts of the representative frame; computing a plurality of spatial average luminance values for angular sectors in concentric rings of the representative frame; normalizing the spatial average luminance values e.g. by the maximum spatial average luminance value; computing the gradient of normalized spatial average luminance values.
In a second preferred embodiment, the method comprises a step of extracting the on-screen picture vertical stability from the digital representation of the copy, the step of extracting the on-screen picture vertical stability further comprising computing a vertical frame offset for each frame of the digital representation of the copy belonging to a stable segment of the copy. According to particularly advantageous characteristics, the step of extracting the vertical stability comprises computing a global statistic on each frame of the digital representation of the copy; extracting shot boundaries to determine a plurality of shots in the digital representation from global statistics; isolating stable segments within each shot comprising at least F frames from global statistics; discarding stable segments whose temporal activity, measured from the global statistics, exceeds a determined threshold; applying a global motion estimator on each frame of each stable segments resulting in the global horizontal and vertical frame offset; selecting N segments with the lowest offset variations; computing an average of variance on the vertical frame offset for each N selected segment; computing a histogram of the vertical frame offset for each N selected segment.
In a third preferred embodiment, the method comprises a step of extracting the temporal illumination pulse from the digital representation of the copy, the step of extracting the temporal illumination pulse further comprising computing an average value along each line of luminance frames of the digital representation of the copy. According to particularly advantageous characteristics, the step of extracting the temporal illumination pulse further comprises applying a first temporal derivative filter on the line luminance average values resulting in the temporal derivative of the luminance average values along the rows of the frame; clipping the time derivative luminance average values to a threshold d; computing a temporal Fourier transform on each clipped time derivative luminance average values; computing the average of all the Fourier transform spectrum magnitudes; applying a low pass filter; extracting the amplitude, width and location of the N most significant peaks of the average of temporal Fourier transform magnitude.
Other characteristics and advantages of the invention will appear through the description of a non-limiting embodiment of the invention, which will be illustrated, with the help of the enclosed drawings.
In the FIGURE, the represented blocks are purely functional entities, which do not necessarily correspond to physically separate entities. These functional entities may be implemented as hardware, software, or a combination of software and hardware; furthermore, they may be implemented in one or more integrated circuits.
Each feature or variant disclosed in the description, in the claims or in the drawings may be provided independently or in any appropriate combination.
In a first step E1, the features F1, F2, F3 are extracted from the copy 100. In a second step E2, the features F1, F2, F3 are normalized and concatenated into a multidimensional numerical vector F. The resulting features vector F is then input to a state-of-the-art supervised classification system in a step E3. The classifier is advantageously previously trained during a learning phase. The learning phase provides both the feature normalization parameters such as the mean and standard deviation for each feature and the classifier decision boundaries. After the classification step E3, the copy 100 of a recaptured video content is either classified as recorded from the digital projection of the video content 101 or classified as recorded from the projection of the film (i.e the celluloid print) of the video content 102. The term “celluloid print” or “film” is equally used in the description. The term “copy”, “digital representation of the copy” or “video sequence” is also equally used in the description.
Hereafter, the rationale for the features choice based on observed artefacts and the feature extraction is detailed for each proposed feature.
Spatial Illumination Uniformity Feature Extraction
The spatial illumination uniformity of digital cinema projectors, as enforced by the Digital Cinema Initiatives specification, is significantly higher than what is attainable with film projectors. Statistics are therefore extracted by capturing such property in the captured video sequence.
All luminance frames Y of the video sequence are first combined to yield a single representative frame Yrep. In a variant, the representative frame could simply be the temporal average of the individual frames of the video. The temporal averaging process is expected to attenuate interferences from the content and the representative frame Yrep should converge toward a flat uniform image in case of uniform lighting. In contrast, in case of non uniform lighting, the representative frame should converge toward an image made of concentric uniform rings centered at the projection center.
In another variant, the representative frame Yrep could be obtained by:
In an optional step, the computed representative frame Yrep is then post-processed to facilitate subsequent feature extraction. A typical post-processing operation is to crop-out parts of the camcorded copy that do not capture projected video content e.g. the top and bottom portions of video corresponding to parts below and above the cinema screen that happen to be captured at the same time. These parts of the representative frame Yrep can be characterized by their lower temporal activities and can be automatically cropped out based on some thresholding operation that considers the temporal variance averages along lines and/or columns.
Finally, discriminating features are extracted from the post-processed representative frame Z in order to capture the spatial uniformity (or the absence thereof) of the illumination. At the moment, the following features have been considered:
There may be some normalization steps involved in the feature extraction process in order to guarantee some calibration across movies. For instance, one could normalize average luminance values by their maximum values.
In a preferred embodiment, the step of extracting the spatial uniformity comprises:
Vertical On-Screen Picture Stability Feature Extraction
In celluloid film projections, it is very common to observe a slight vertical instability. This instability is due to the film sprocket-hole mechanism and manifests itself as a pseudo-periodic, mostly vertical, sub-pixel inter-frame global motion. Statistics are therefore extracted by capturing such property in the captured video sequence.
Since global motion estimators perform better on video sequence with low motion activity, it is common practice to isolate segments of the analyzed video sequence where there is very little activity. To do so, the video is first inspected using global statistics in order to pinpoint strong discontinuities indicating shot boundaries. Known global statistics are for instance the frame average luminance value, the frame luminance histogram, the frame RASH descriptor as detailed in “A video fingerprint based on visual digest and local fingerprint” from A. Massoudi et al. in IEEE international Conference on Image Processing. Next, within each shot, the f-frames long segment (e.g. f=50) with low temporal activity according to the considered global statistic is retained. This could be f frames surrounding the most stable frames in the shot. Alternatively, it could also be the segment of f frames whose overall activity is the lowest in the shot.
At this stage, the collection of retained low-activity segments can be further trimmed down by discarding all segments whose activity, as measured by the considered global statistic, exceeds some pre-defined threshold Tactivity.
Each f frames-long remaining segment S is then fed into a global motion estimator, for instance using a phase correlation algorithm, which returns, for each frame of the segment, a global vertical and horizontal offset with reference to one selected frame in S.
This estimation could be corrupted when the camcorder was moved during the capture, for instance when the camcorder is handheld. To alleviate this issue, segments for which “large” global motion is detected can be discarded. Indeed the pseudo-periodic jitter induced by the deteriorated sprocket-holes is expected to yield very low amplitude vertical motion and any segment whose global motion exceeds at some point a threshold Tmotiong could be discarded. Alternatively, one could select the Nmotion segments with the lowest vertical or horizontal motion considering the maximum amplitude or the variance.
After this last clean-up operation, features are derived from the f-dimensional vertical motion vectors extracted from the retained low activity segments. According to different variants, the following features have been considered:
In a preferred embodiment, the step of extracting the on-screen picture vertical stability comprises:
Temporal Illumination Pulse Feature Extraction
During projection, a camcorder and a projector are never fully synchronized temporally. Firstly, the acquisition frequency and the projection frequency always differ by at least a small amount. Secondly, and possibly more critically, there is also a phase difference between the camcorder and the projector that is to say their clocks are not aligned. As a result, the frames of the camcorded video sequence partly integrate the period of darkness created by the shutter of the projector in-between two successive projected frames. This frequency difference creates a periodical global luminance variation. Since the period of darkness when the shutter interrupts the light in a film projector (during film advance between two successive frames) is much longer compared to the corresponding period in a digital projector, the resulting pulse phenomenon is more important in captures of film projections.
Depending on the intrinsic parameters of the camera (sensor type, acquisition frequency, etc), the above mentioned temporal illumination pulse may reveal itself in different forms for instance a global luminance pumping effect or a large dark horizontal band going down. In any case, it should introduce some statistical artifacts along the time axis.
In a first step, the luminance frames of the video are averaged along the rows of the frames. As a result, the video can be represented as a R×T matrix M, where R is the number of pixel rows in the video sequence and T the number of frames in the video sequence. In this representation, a global luminance pumping effect would be revealed by regular vertical dark bands. Similarly, a dark horizontal band moving down the screen would be revealed by slightly inclined dark bands in this representation.
In an optional step, the matrix M is then post-processed to facilitate subsequent feature extraction. A typical post-processing operation is to crop-out parts of the camcorded copy that do not capture projected video content e.g. the top and bottom portions of video corresponding to parts below and above the cinema screen that happen to be captured at the same time. These parts of the matrix M can be characterized by their lower temporal activities and can be automatically cropped out based on some thresholding operation that considers the temporal variance along rows.
To diminish the impact of host interference, a first derivative filter is applied to M to obtain the time derivative D of the frames averaged along the rows.
In order to attenuate undesired artifacts at shot boundaries, the values of D are clipped at a preset value d. In other words, all values of D whose absolute values exceed d are set to ±d.
At this stage, a Fourier transform is applied to the rows of D in order to reveal unexpected frequencies in the spectrum introduced by the temporal illumination pulse. It should be noted that enforcing this row by row strategy permits to alleviate any issue related to possible frame interleaving. In order to get more stable statistics, the magnitude of the spectrums are then averaged along the columns in order to obtain the average spectrum magnitude of the time derivative of the rows average luminance. Again, it should be noted that averaging magnitude instead of the full spectrums allows to get rid of any problem due to possible phase differences between rows. The resulting average spectrum is subsequently low-pass filtered.
Features are then derived from this low-pass average spectrum. In different variants, the following features have been considered:
In a preferred embodiment, the step of extracting the temporal illumination pulse comprises:
Naturally, the invention is not limited to the embodiments previously described.
In particular, the invention is compatible with any other application than forensic analysis which needs to distinguishing digital copies of a movie either issued from a digital projection or a celluloid film projection.
| Number | Date | Country | Kind |
|---|---|---|---|
| 11305899.4 | Jul 2011 | EP | regional |
