Neural network stylization of watermark signals
AssigneeDigimarc Corporation
InventorsTomas Filler, Nicholas Anderson, Ajith M Kamath
Life of this patent
Estimating the term…
Abstract
A watermark image is stylized using a neural network. Stylization can employ a composite loss function in which a content loss term is based on image similarity to the watermark image itself, rather than based on similarity to layer activations produced by the neural network. The composite loss function may further avoid use of mean-squared error in connection with content loss, to allow the network freedom to produce a stylized image having a mean luma value that is remote from a mean luma value of the watermark image. An exemplary application generates texture patterns that simulate leather, yet convey plural-symbol payloads. Many other features and arrangements are also detailed.
Claims12
A watermarked image embedded in a physical surface or stored on a non-transitory computer readable medium, previously derived from an original image, the watermarked image having a mean value that is not between 124 and 131, wherein at least one of the following conditions is satisfied:
- (a) a gradient between a pair of adjoining pixels in the watermarked image differs, by more than 30, from a gradient between a spatially-corresponding pair of pixels in the original image;
- or
- (b) a pixel value in the watermarked image differs, by more than 64, from a pixel value in the original image;
- wherein the watermarked image is structurally different than the original image, with edges having been both added to and removed from the original image.
Dependencies are inferred from the claim text.