Granted Patent B2

US 12,731,428 · App. 19/340,465 · Granted Sep 8, 2026

Inventors: Soumitri Naga Kolavennu (Plymouth, MN); Priyanka Singhal (Atlanta, GA); Varshini Sriram (Menlo Park, CA); Lavanya Basavaraju (Minneapolis, MN)

Assignee: U.S. Bank National Association

G06V30/42G06N20/00G06Q20/4016G06Q40/02G06V30/414G06V30/418G06V40/33G06Q20/042

Abstract

Claims (107)

1 . A method for detecting fraudulent documents, the method comprising:

receiving a document image associated with a document;

processing the document image to generate a processed document image for fraud detection, including:

converting the document image to a format the same as reference documents to avoid, during an object detection processing, object detection issues caused by a format difference between the reference documents and the document image, and

resizing the document image to predetermined lower resolution dimensions to avoid object detection issues during the object detection processing;

detecting a plurality of document objects of interest on the processed document image using the object detection processing;

determining coordinates of the plurality of document objects of interest in the processed document image;

generating one or more document bounding boxes enclosing one or more of the plurality of document objects of interest;

reconverting the processed document image back into original resolution dimensions after the one or more bounding boxes are generated to locate the objects of interest in the document image;

calculating Intersection over Union (IoU) metrics for the one or more document bounding boxes and one or more reference document bounding boxes enclosing one or more corresponding reference document objects of interest on a reference document, wherein the calculating is based on the one or more document bounding boxes on the reconverted processed document image to enable accurate comparison of the objects of interest for fraud detection;

determining an IoU score based on the IoU metrics;

determining a fraud score based on the IoU score; and

either:

determining the document is valid based on the fraud score and approving the document; or

determining the document exhibits an indication of invalidity based on the fraud score and flagging the document.

2 . The method of claim 1 , further comprising:

determining one or more of the plurality of document objects of interest are variable objects of interest in the document image;

creating a document stock image that excludes the variable objects of interest;

comparing the document stock image and a reference document stock image using an algorithm; and

determining an image pattern score based on the comparison of the document stock image and the reference document stock image,

wherein determining the fraud score is based on the image pattern score.

3 . The method of claim 2 , wherein the algorithm:

creates hashes in a manner that have close Euclidean distance or cosine similarity between documents from the same document stock and a large Euclidean distance or cosine similarity for documents from different document stocks or altered document stock.

4 . The method of claim 3 , wherein comparing the document stock image and the reference document stock image using the algorithm includes:

using the algorithm to create and store a target hash based on the reference document stock image;

using the algorithm to create a hash based on the document stock image; and

calculating any one of (i) Euclidean distances between the target hash and the hash, (ii) a cosine similarity between the target hash and the hash, or (iii) a combination of (i) and (ii).

5 . The method of claim 1 , further comprising:

detecting a signature of the document;

comparing the signature of the document and one or more reference signatures using an algorithm; and

determining a signature score based on the comparison of the signature of the document and the one or more reference signatures, wherein determining the fraud score is based on the signature score.

6 . The method of claim 5 , wherein comparing the signature of the document and the one or more reference signatures using the algorithm includes:

creating a signature hash model with a target signature data point;

populating the signature hash model with data points representing the signature of the document and the one or more reference signatures;

determining the distances of the data points representing the signature of the document and the one or more reference signatures from the target signature data point;

determining a target distance from the target signature data point based on the distances of the data points representing the one or more reference signatures from the target signature data point;

comparing the distance of the data point representing the signature of the document and the target distance; and

determining a signature similarity of the signature of the document based on the comparison of the distance of the data point representing the signature of the document and the target distance.

7 . The method of claim 6 , wherein comparing the signature of the document and the one or more reference signatures includes:

using the algorithm to create and store a target hash based on the one or more reference signatures;

using the algorithm to create a hash based on the signature of the document; and

calculating any one of (i) Euclidean distances between the target hash and the hash, (ii) a cosine similarity between the target hash and the hash, or (iii) a combination of (i) and (ii).

8 . A method for detecting document fraud, the method comprising:

receiving a document image associated with a document;

processing the document image to generate a processed document image for fraud detection, including:

converting the document image to a format the same as reference documents to avoid, during an object detection processing, object detection issues caused by a format difference between the reference documents and the document image, and

resizing the document image to predetermined lower resolution dimensions to avoid object detection issues during the object detection processing;

detecting a plurality of document objects of interest on the document image using the object detection processing;

determining one or more of the plurality of document objects of interest are variable objects of interest in the document image;

creating a document stock image based on the one or more of the plurality of document objects of interest, wherein the document stock image excludes the variable objects of interest;

generating one or more document bounding boxes enclosing one or more of the plurality of document objects of interest;

reconverting the processed document image back into original resolution dimensions after the one or more bounding boxes are generated to locate the objects of interest in the document image;

calculating Intersection over Union (IoU) metrics for the one or more document bounding boxes and one or more reference document bounding boxes enclosing one or more corresponding reference document objects of interest on a reference document;

determining an IoU score based on the IoU metrics, wherein the calculating is based on the one or more document bounding boxes on the reconverted processed document image to enable accurate comparison of the objects of interest for fraud detection;

comparing the document stock image and a reference document stock image;

determining an image pattern score based on the comparison of the document stock image and the reference document stock image; and

determining a fraud score based on (1) the IoU score and (2) the image pattern score.

9 . The method of claim 8 , further comprising:

determining the document is valid based on the fraud score; and

approving the document.

10 . The method of claim 8 , further comprising saving the document as a new reference document.

11 . The method of claim 8 , further comprising:

determining the document requires review based on the fraud score; and

flagging the document.

12 . The method of claim 8 , wherein comparing the document stock image and the reference document stock image includes:

creating hashes of the document stock image and the reference document in a manner that have close Euclidean distance or cosine similarity between documents from the same document stock and a large Euclidean distance or cosine similarity for documents from different document stocks or altered document stock.

13 . The method of claim 8 , wherein comparing the document stock image and the reference document stock image includes:

creating and storing a target hash based on the reference document stock image;

creating a hash based on the document stock image; and

calculating any one of (i) Euclidean distances between the target hash and the hash, (ii) a cosine similarity between the target hash and the hash, or (iii) a combination of (i) and (ii).

14 . A method for detecting document fraud, the method comprising:

receiving a document image associated with a document;

processing the document image to generate a processed document image for fraud detection, including:

converting the document image to a format the same as reference documents to avoid, during an object detection processing, object detection issues caused by a format difference between the reference documents and the document image, and

resizing the document image to predetermined lower resolution dimensions to avoid object detection issues during the object detection processing;

detecting a plurality of document objects of interest on the processed document image using the object detection processing;

generating one or more document bounding boxes enclosing one or more of the plurality of document objects of interest;

reconverting the processed document image back into original resolution dimensions after the one or more bounding boxes are generated to locate the objects of interest in the document image;

calculating Intersection over Union (IoU) metrics for the one or more document bounding boxes and one or more reference document bounding boxes enclosing one or more corresponding reference document objects of interest on a reference document, wherein the calculating is based on the one or more document bounding boxes on the reconverted processed document image to enable accurate comparison of the objects of interest for fraud detection;

determining an IoU score based on the IoU metrics;

detecting a signature of the document;

comparing the signature of the document and one or more reference signatures;

determining a signature score based on the comparison of the signature of the document and the one or more reference signatures; and

determining a fraud score based on (1) the IoU score and (2) the signature score.

15 . The method of claim 14 , further comprising:

determining the document is valid based on the fraud score and approving the document; or

determining that the document exhibits an indication of invalidity based on the fraud score and flagging the document.

16 . The method of claim 14 , wherein reconverting the processed document image back into the original resolution dimensions comprises:

determining coordinates of the plurality of document objects of interest in the processed document image; and

mapping the one or more document bounding boxes based on the coordinates.

17 . The method of claim 14 , further comprising:

determining one or more of the plurality of document objects of interest are variable objects of interest in the document image;

creating a document stock image, wherein the document stock image excludes the variable objects of interest;

comparing the document stock image and a reference document stock image; and

determining an image pattern score based on the comparison of the document stock image and the reference document stock image, wherein determining the fraud score is based on the image pattern score.

18 . The method of claim 17 , wherein comparing the document stock image and the reference document stock image includes:

creating a target hash based on the reference document stock image;

creating a hash based on the document stock image; and

calculating any one of (i) Euclidean distances between the target hash and the hash, (ii) a cosine similarity between the target hash and the hash, or (iii) a combination of (i) and (ii).

19 . The method of claim 14 , wherein comparing the signature of the document and the one or more reference signatures includes:

creating a signature hash model with a target signature data point;

populating the signature hash model with data points representing the signature of the document and the one or more reference signatures;

determining the distances of the data points representing the signature of the document and the one or more reference signatures from the target signature data point;

determining a target distance from the target signature data point based on the distances of the data points representing the one or more reference signatures from the target signature data point;

comparing the distance of the data point representing the signature of the document and the target distance; and

determining a signature similarity of the signature of the document based on the comparison of the distance of the data point representing the signature of the document and the target distance.

20 . The method of claim 14 , further comprising saving the document as a new reference document.

Assignments (1)

Continuity (2)

References Cited (36)