InvisibleImageWatermarker<T>
Models & Types · AiDotNet.Safety.Watermarking
Keyed spread-spectrum image watermarker that hides a pseudo-random +/-1 sequence in the image’s mid- and high-frequency block DCT coefficients and detects it by correlation.
Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → ImageSafetyModuleBase<T> → ImageWatermarkerBase<T> → InvisibleImageWatermarker<T>
Known derived types: ImageWatermarker<T>
For Beginners
the image is rewritten as a sum of wave patterns (the DCT). A secret, random-looking pattern of tiny nudges is added to many of those waves at once. No single nudge is visible, but someone with the key can add up all the nudges and see a clear signal; without the key the nudges look like noise.
How It Works
Builds on the spread-spectrum idea of Cox, Kilian, Leighton and Shamoon (1997), spreading a key-derived sequence
over many DCT coefficients so each changes imperceptibly, but it is not their scheme. Cox et al. take a full-frame
DCT, embed multiplicatively (c’ = c * (1 + a * x)) in the perceptually most significant, mostly low-frequency
coefficients, and detect by comparing against the original image. This class is a blind additive variant: the
image is cut into independent tiles of at most MaxTileSize pixels per side,
each tile gets its own DCT and its own key-derived sequence, the sequence is added (c’ = c + a * x), and detection
needs only the key, not the original. The coefficients used in each tile are those at normalized frequency radius
0.1-0.8, excluding the first row and column, which carry the energy of tile borders and straight gradients.
Constructors
| Constructor | Summary |
|---|---|
InvisibleImageWatermarker(double watermarkStrength, string key) |
Initializes a new spread-spectrum DCT image watermarker. |
Properties
| Property | Summary |
|---|---|
string ModuleName { get; } |