Artwork generated to convey digital messages, and methods/apparatuses for generating such artwork
Features from a style image are adapted to express a machine-readable code. For example, grains of rice depicted in a style image may be positioned to create a pattern mimicking that of a machine-readable code. The resulting output image can then be used as a graphical component in product packaging (e.g., as a background, border, or pattern fill), while also serving to convey a product identifier to a compliant reader device (e.g., a retail point-of-sale terminal). In some embodiments, a neural network is trained to apply a particular style image to machine readable codes. A great variety of other features and arrangements are also detailed.
1 . A method comprising the acts:
digitally-watermarking a first image to encode a watermark pattern including a plural-bit payload therein, yielding a second image; and
neural network processing the second image to yield a neural network processed second image, the neural network processing to: i) reduce a difference measure between the neural network processed second image and a style image, or ii) increase a similarity measure between the neural network processed second image and the style image, wherein the neural network has been pre-trained by applying back propagation to adjust filter weights in the neural network according to a first loss function and a second loss function, the first loss function corresponding to the difference measure or similarity measure between a processed input image and a style image, and the second loss function corresponds to differences between a watermark pattern in the processed input image and an input watermark pattern.
2 . The method of claim 1 in which the difference measure is based on a difference between a Gram matrix for the neural network processed second image and a Gram matrix for the style image.
3 . The method of claim 2 that includes incorporating some or all of the neural network processed second image within label artwork for food product packaging.
4 . The method of claim 1 in which the plural-bit payload is recoverable by a digital watermark decoder from the neural network processed second image.
5 . A method for producing a code for message signaling through an image, the method employing a deep neural network, the deep neural network having an input, one or more outputs, and plural intermediate layers, each layer of the plural intermediate layers comprising plural filters, each filter of the plural filters characterized by plural parameters that define a response of the each filter to a given input, the method comprising:
iteratively adjusting an input test image, based on results produced by the plural filters of said neural network, until the input test image adopts both (1) style features from one image, and (2) signal encoding features from a second image; and
outputting an iteratively adjusted input test image, the iteratively adjusted input test image comprising signal-encoding features, in which the signal-encoding features are machine-readable from the iteratively adjusted input test image.
6 . A method comprising:
training a neural network, including: applying back propagation to adjust filter weights in the neural network according to a first loss function and a second loss functions, the first loss function corresponding to a difference measure or similarity measure between a digitally-watermarked input image and an input style image to i) reduce a difference measure between the digitally-watermarked input image and the input style image, or ii) increase a similarity measure between the digitally-watermarked input image and the input style image, and the second loss function corresponding to differences between a watermark pattern in an image representing a stylized version of the digitally-watermarked input image, and an input watermark pattern, said training yielding a trained neural network;
digitally-watermarking a first image to encode a watermark pattern including a plural-bit payload, thereby yielding a second image;
providing the second image and a style image to the trained neural network; and
processing, with the trained neural network, the second image in accordance with the style image to yield a stylized second image comprising the watermark pattern.
7 . The method of claim 6 wherein the style image comprises artwork.
8 . The method of claim 7 wherein the artwork comprises packaging artwork, and the stylized second image is incorporated into packaging, the plural-bit payload being machine readable from an image captured of the packaging using digital watermark decoding.
9 . The method of claim 6 wherein the style image comprises a photograph.
10 . A method comprising:
receiving a pattern that encodes a plural-symbol payload; and
applying a style to said pattern, using a previously-trained neural network, to produce a stylized output image, the style being based on a style image having various features, wherein the previously-trained neural network adapts features from the style image to express details of the received pattern, to thereby produce an image in which features from the style image contribute to encoding of said plural-symbol payload, wherein the previously-trained neural network is trained to incorporate a digital watermark pattern that encodes a payload into artwork to produce the stylized output image by applying back propagation to adjust filter weights in a neural network according to a first loss function and a second loss function, the first loss function corresponding to a similarity measure between the artwork and artwork incorporating the digital watermark pattern, and the second loss function corresponding to differences between a digital watermark pattern in the artwork and the digital watermark pattern that encodes the payload.
11 . The method of claim 10 wherein the style image comprises a photograph, and the previously-trained neural network incorporates the digital watermark pattern that encodes the payload into the photograph or an image derived from the photograph.
12 . The method of claim 10 wherein the previously-trained neural network is trained using a dataset comprising digital watermark patterns and corresponding style images.
13 . The method of claim 10 wherein the artwork comprises packaging artwork, and the stylized output image is incorporated into packaging, the plural-symbol payload being machine readable from an image captured of the packaging using digital watermark decoding.
14 . The method of claim 10 wherein the artwork comprises a digital image, and the stylized output image comprises the digital image embedded with the digital watermark, the plural-symbol payload being machine readable from the digital image embedded with the digital watermark.
15 . The method of claim 10 wherein the previously-trained neural network comprises a deep neural network having an input, one or more outputs, and plural intermediate layers, each layer of the plural intermediate layers comprising plural filters, each filter of the plural filters characterized by plural parameters that define a response of the each filter to a given input, the deep neural network being trained according to a training method comprising the acts:
iteratively adjusting an input test image, based on results produced by the plural filters of the deep neural network, until the input test image adopts both (1) style features from one image, and (2) signal-encoding features from a second image; and outputting an iteratively adjusted input test image, the iteratively adjusted input test image comprising the signal-encoding features, in which the signal-encoding features are machine-readable from the iteratively adjusted input test image.
16 . The method of claim 5 comprising a training process including the iteratively adjusting of the input test image by adjusting filters of the deep neural network through backpropagation to adjust filter weights of layers in the deep neural network according to a first loss function and a second loss function, the first loss function corresponding to a similarity measure between the one image and the one image incorporating a digital watermark, and the second loss function corresponding to a differences between the digital watermark in the one image incorporating the digital watermark and a reference digital watermark.
17 . The method of claim 15 , wherein the signal-encoding features comprise the digital watermark, which is machine readable with a digital watermark decoder.
18 . The method of claim 10 , wherein the previously-trained neural network comprising a feed-forward neural network trained using a perceptual loss function that includes:
(i) a content loss based on feature activations of a convolutional neural network, and
(ii) a style loss based on correlations between feature maps of the convolutional neural network.