SysCoP is a set of methods to embed robust labels into different types of images. The system supports gray-scaled, color, and binary images. These methods share an algorithm framework for both label writing and reading.

The framework as shown in Figure 1 for label writing and Figure 2 for label reading is composed of two steps. The first step generates a pseudo random position sequence for selecting blocks where the code is embedded. This step is denoted as a function T_s(y, U_k) where y is the image data to be labeled, and U_k is the user-supplied secret key. The second step embeds or retrieves the code into or from the blocks specified in the position sequence. The methods for embedding or reading code depend on types of images, and will be next described individually.

Figure 1: Scheme for embedding digital watermarks into image data.

Figure 2: Recovery of digital watermark from image data.

The function T_s(y,U_k) first extracts some features from the image data and then use them together with the user secret key as the seeds for position sequence generation. Ideally, the features of the image data used here must meet the following requirements:

• Must be robust against simple image processing that does not affect the visual quality of the image.
• Must be image-dependent, i.e. the image can be identified by these features extracted from the image data.

These requirements are to some extent contradictory and must be balanced carefully.

The second step is the actual embedding of the position sequence:.

• The generated sequence is used to create a sequence of pixel-mapped locations indicating where the code is to be embedded
• The blocks of 2D image y(k,l) where k,l are the indices of discrete image points are locally transformed and quantized into the frequency domain at the locations selected in the previous step in a manner reflecting acceptable information (quality) loss in the image for the application to produce a 2D image residual n(k,l) in which the mark will actually be embedded.
• The code values (1 or 0) representing the binary digits of the code being embedded are superimposed on the signal n(k,l) at the selected locations.
• The quantized data is decoded and inversely transformed back into the spatial domain yielding the marked image.

The transformations used must be appropriate for the given type of image (color, greyscale, bitmap) and must also select the data to be manipulated in such a way as to balance robustness and quality degradation.

One possibility for such a transformation is using the DCT used in the JPEG compression algorithm.

Images transformed into the frequency domain have the majority of information at low frequencies, high frequencies have less visual impact. The concentration of spectral energy

is in the lower frequency coefficients

where |k,l| is the standard Euclidean norm of the vector (k,l). The spectral power is also expected to be a nonincreasing function of the normal of the frequency vector

Due to the observation above it is permissible to use a coarser quantization on the higher frequency coefficients than on lower frequency coefficients; this is used by the JPEG transform quantizer by increasing qs(k,l) as a function of the increasing frequency vector normal.

Using these assumptions several signals can be derived from the image data Y(i,j) which meet the requirements for embedding the watermark signal.

Figure 3: Example of marked and unmarked images
(photos courtesy of NASA)

Original	Marked

It should be noted that the algorithm is robust against the distortions and quality losses induced by digital/analog (printing) and subsequent analog/digital (scanning) conversions, provided the two conversions occur at a sufficient quality level.