Back to results
US 11,909,933 B1Granted
One-shot neural network for stamp detection
AssigneeKYOCERA Document Solutions Inc.
InventorDongpei Su
Life of this patent
Estimating the term…
Abstract
A method of detecting a digital stamp pattern involves operating a scanning device to create a page image scanned from a document page, inputting the page image into a one-shot trained neural network, the one-shot trained neural network configured to recognize a copy-guard digital stamp pattern using one-shot learning, analyzing the page image using the one-shot trained neural network to detect the copy-guard digital stamp pattern, and on condition the copy-guard digital stamp pattern is detected, issuing an electronic alert.
Claims18
A method of detecting a digital stamp pattern, the method comprising:
- operating a scanning device to create a page image scanned from a document page;
- inputting the page image into a one-shot trained neural network, the one-shot trained neural network configured to recognize a copy-guard digital stamp pattern using one-shot learning;
- analyzing the page image using the one-shot trained neural network to detect the copy-guard digital stamp pattern;
- and
- on condition the copy-guard digital stamp pattern is detected, issuing an electronic alert;
- wherein configuring the one-shot trained neural network to recognize the copy-guard digital stamp pattern using the one-shot learning comprises:
- generating a set of training page images, wherein each of the training page images of the set of training page images comprises one of:
- a background page;
- and
- the background page and at least one digital stamp pattern, the at least one digital stamp pattern being one of the copy-guard digital stamp pattern and another digital stamp pattern;
- grouping the training page images into image pairs, wherein the image pairs having matching digital stamp patterns are labeled positive image pairs and the image pairs having different digital stamp patterns are labeled negative image pairs;
- configuring the synthesized images with print and copy noise by printing then scanning the synthesized images;
- inputting the image pairs into a twin neural network;
- applying a contrastive loss function to an output of the twin neural network to calculate loss values using both images of the image pairs;
- updating model weights using the loss values to configure a refined twin neural network;
- mapping the refined twin neural network to layers supported by a deployment target;
- translating the mapped refined twin neural network to a fixed point model neural network;
- and
- deploying the fixed point model neural network on the deployment target as the one-shot trained neural network.
16 more elements
A computing apparatus comprising:
- a processor;
- and
- a memory storing instructions that, when executed by the processor, configure the apparatus to:
- operate a scanning device to create a page image scanned from a document page;
- input the page image into a one-shot trained neural network, the one-shot trained neural network configured to recognize a copy-guard digital stamp pattern using one-shot learning;
- analyze the page image using the one-shot trained neural network to detect the copy-guard digital stamp pattern;
- and
- on condition the copy-guard digital stamp pattern is detected, issue an electronic alert;
- wherein configuring the one-shot trained neural network to recognize the copy-guard digital stamp pattern using the one-shot learning comprises:
- generate a set of training page images, wherein each of the training page images of the set of training page images comprises one of:
- a background page;
- and
- the background page and at least one digital stamp pattern, the at least one digital stamp pattern being one of the copy-guard digital stamp pattern and another digital stamp pattern;
- group the training page images into image pairs, wherein the image pairs having matching digital stamp patterns are labeled positive image pairs and the image pairs having different digital stamp patterns are labeled negative image pairs;
- configure the synthesized images with print and copy noise by printing then scanning the synthesized images;
- input the image pairs into a twin neural network;
- apply a contrastive loss function to an output of the twin neural network to calculate loss values using both images of the image pairs;
- update model weights using the loss values to configure a refined twin neural network;
- map the refined twin neural network to layers supported by a deployment target;
- translate the mapped refined twin neural network to a fixed point model neural network;
- and
- deploy the fixed point model neural network on the deployment target as the one-shot trained neural network.
19 more elements
Dependencies are inferred from the claim text.