Abstract
(Englisch)
|
Project context
Digital media have become common and have extended the applications of traditional analog media. Moreover, the popularity of the Internet has clearly demonstrated the commercial potential of the digital multimedia market and consumers are investing heavily in digital audio, image and video recorders and players. Unfortunately, digital networks and multimedia also afford virtually unprecedented opportunities to pirate copyrighted material. Therefore, copyright enforcement for images and videos based on digital watermarking are proposed as a method for discouraging illicit copying and distribution of copyrighted documents.
In the JEDI-FIRE project, the Computer Vision Group, University of Geneva (associated partner to DCT - Digital Copyright Technologies Ltd., Switzerland) concentrates on methods for watermarking digital images and videos; DCT Switzerland concentrates on the development of a security architecture for copyright creation and enforcement.
Requirements for digital watermarking
A digital watermark is an invisible signature that is hidden in an image or a video; this signature, usually of length 60 to 80 bits, typically contains information regarding the owner of the document as well as a time stamp. A secret key is necessary to embed a watermark in an image. Two types of watermarks exist, public and private. A public watermark is decodable by any person using the appropriate software, without the need for the secret key; this allows to determine whether or not an image is copyrighted. A secret watermark contains information that can only be read by someone having the secret key. In both cases, only the key owner can remove the mark.
In order to provide for public acceptance as well as for efficient copyright enforcement, the watermark has to satisfy a number of constraints:
· it has to be perceptually invisible;
· it must be possible to recover the watermark without having the original, unmarked images (oblivious watermarking);
· it has to be resistant to common image processing operations, such as photometric transformations, scanning and reprinting, JPEG and MPEG compression, geometric transformations (such as cropping, translation, rotation, scaling, aspect ratio change, flipping, general linear or affine transformation, etc.), regular removal of rows/columns or video frames, video frame rate change (temporal scaling), etc.;
· it has to be resistant to malicious attacks, such as based on signal processing or cryptographic techniques.
Watermarking approaches
In the course of this project we have investigated a number of approaches to the watermarking of images and videos. We summarize below some basic principles of these approaches; a full presentation can be found in the references. The video watermarking techniques we have developed are often very similar to the techniques used for marking images, so they are not specifically described here.
Almost all approaches have in common the fact that the watermark string itself (e.g. 60 to 80 bits) is transformed into a spread spectrum sequence, in order to provide for better noise immunity. The values of this sequence are then used to modify either the values of the image (or video) to be marked, or the values of the image after suitable transformation (e.g. DCT, Fourier, wavelets). All methods must also give some means of recovering from geometrical transformations that the image (or video) has incurred.
The first approach we have investigated was characterized by (see Figure 1):
· the insertion of the spread spectrum sequence in the Fourier or DCT domain;
· the use of a template to recover from geometric transformations. The template is a key-based grid, inserted in the transform domain, and detected in a log-log or log-polar mapping of this transform domain. The purpose of this mapping is to transform an aspect ratio change, or a rotation/scaling operation, into a translation. After the mapping, the template that is found is correlated with a known, reference template, and the correlation peak indicates the parameters of the geometrical transformation undergone by the image. It is then possible to compensate for the geometric transformation, and subsequently to decode the watermark. A similar idea was applied for videos (see Figure 2).
Amongst the subsequent improvements made to this approach, the main ones have been:
· the use of perceptual masks based on the study of noise visibility functions, in order to decrease the visibility of the watermark without altering its robustness. The basic idea here is that the mark should be hidden in regions of the image or video where it is less visible, that is edge regions (as opposed to flat regions);
· the use of point matching algorithms in order to extract the geometrical transformation undergone by an image without having the need for the log-log or log-polar mappings mentioned above. This allowed to have watermarks resistant to general affine transformations;
· the use of a Bayesian test to assess the presence of watermark, even if the watermark itself was not recoverable due to distortions or attacks.
More recently we have been interested in tackling the problem of recovering from geometric distortions from a more general viewpoint. To this effect, we have started to develop a families of methods that consider the watermark as a spread-spectrum periodical sequence that can be inserted in either of the original image domain, or a transformed domain (Fourier, wavelets). This alleviates the need for using a template: the periodical features of the watermark itself, observable by autocorrelation, In order to better understand the weaknesses of our methods, and consequently to develop more robust ones, we have also devoted a significant research effort to the topic of watermark attacks. We have developed a new family of sophisticated stochastic watermark attacks, that take into account the statistics of the image and of the watermark. These attacks are currently able to hide the watermark and render it undetectable for all academic and commercial softwares we know of, this without altering the quality of the image.
Future prospects
It is worth mentioning that starting around the end of 1998, the research team (CUI - University of Geneva, DCT - Zürich) has decided to team up with the Signal Processing Laboratory of the EPFL (Profs. Murat Kunt, Touradj Ebrahimi). The current 'Swiss watermarking task force' has now established itself as one of the most prominent one, worldwide.
We will at the same time pursue our basic research efforts and the commercialization enterprise. Regarding the basic research efforts, some of the directions are:
· autocorrelation-based watermarking: to alleviate the need for explicit template insertion and retrieval;
· use of other coding principles, such as Error Correction Codes;
· hierarchical wavelet-based watermarking: to permit greater robustness, and to be compliant with future image standards such as JPEG 2000;
· watermark attacks;
· study of the relationship between watermark capacity, robustness, and visibility;
· watermarking of other types of media, such as MPEG4 videos, 3D objects, as well as watermarking of analog data (secure papers).
re used as a template to recover from geometric transformations (see Figure 3).
|
