Identifying and localizing editorial changes to images utilizing deep learning

Patent metadata

Field	Value
Publication number	US-12505565-B2
Application number	US-202217804376-A
Country	US
Kind code	B2
Filing date	May 27, 2022
Priority date	May 27, 2022
Publication date	Dec 23, 2025
Grant date	Dec 23, 2025

Abstract

The present disclosure relates to systems, methods, and non-transitory computer readable media that utilize deep learning to identify regions of an image that have been editorially modified. For example, the image comparison system includes a deep image comparator model that compares a pair of images and localizes regions that have been editorially manipulated relative to an original or trusted image. More specifically, the deep image comparator model generates and surfaces visual indications of the location of such editorial changes on the modified image. The deep image comparator model is robust and ignores discrepancies due to benign image transformations that commonly occur during electronic image distribution. The image comparison system optionally includes an image retrieval model utilizes a visual search embedding that is robust to minor manipulations or benign modifications of images. The image retrieval model utilizes a visual search embedding for an image to robustly identify near duplicate images.

First claim

What is claimed is:

1 . A non-transitory computer readable medium comprising instructions that, when executed by at least one processor, cause the at least one processor to perform operations comprising:
aligning a first image with a second image to generate an aligned first image; generating a fused feature vector by combining deep features from the aligned first image and the second image; generating a classification for modifications of the first image relative to the second image as benign, editorial, or a different image; and generating one or more visual indicators, from the fused feature vector utilizing one or more neural network layers, identifying locations of editorial modifications in the first image relative to the second image.

2 . The non-transitory computer readable medium of claim 1 , wherein aligning the first image with the second image comprises:
generating an optical flow between the first image and the second image; and warping the first image utilizing a de-warping unit based on the optical flow to generate the aligned first image.

3 . The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
generating a first set of deep features for the aligned first image utilizing a neural network feature extractor; and generating a second set of deep features for the second image utilizing the neural network feature extractor.

4 . The non-transitory computer readable medium of claim 3 , wherein generating the fused feature vector by combining deep features from the aligned first image and the second image comprises:
generating a combination of the first set of deep features and the second set of deep features; and generating the fused feature vector from the combination utilizing a neural network encoder.

5 . The non-transitory computer readable medium of claim 1 , wherein generating the one or more visual indicators, from the fused feature vector utilizing the one or more neural network layers, identifying locations of editorial modifications in the first image relative to the second image comprises:
generating a heat map from the fused feature vector utilizing a multilayer perceptron; and overlaying the one or more visual indicators on the first image based on the heat map.

6 . The non-transitory computer readable medium of claim 1 , wherein operations further comprise generating a database of trusted digital images, wherein the second image is a trusted digital image from the database of trusted digital images.

7 . The non-transitory computer readable medium of claim 1 , wherein generating the classification for the modifications of the first image comprises determining:
a probability that the first image has benign changes; a probability that the first image has editorial changes; and a probability that the first image is a different image from the second image.

8 . The non-transitory computer readable medium of claim 1 , wherein generating the classification for the modifications of the first image comprises generating the classification from the fused feature vector utilizing one or more additional neural network layers.

9 . A system comprising:
one or more memory devices comprising a set of trusted digital images; and one or more processors that are configured to cause the system to: search the set of trusted digital images for a trusted near-duplicate image to a query image; align the query image to the trusted near-duplicate image to generate an aligned query image; generating a fused feature vector by combining deep features from the aligned query image and the trusted near-duplicate image; and determine whether changes to the query image relative to the trusted near-duplicate image comprise benign changes or editorial changes by generating a classification from the fused feature vector utilizing one or more neural network layers.

10 . The system of claim 9 , wherein the one or more processors are further configured to cause the system to generate visual search embeddings for images of the set of trusted digital images and the query image utilizing an image retrieval model.

11 . The system of claim 10 , wherein the one or more processors are further configured to cause the system to learn parameters of the image retrieval model utilizing a contrastive loss.

12 . The system of claim 10 , wherein the one or more processors are further configured to cause the system to generate the visual search embeddings by encoding the images of the set of trusted digital images and the query image utilizing a convolutional neural network that is robust to benign image transformations.

13 . The system of claim 10 , wherein the one or more processors are further configured to cause the system to generate binary hashes of the visual search embeddings.

14 . The system of claim 13 , wherein the one or more processors are configured to cause the system to search the set of trusted digital images for the trusted near-duplicate image to the query image by comparing a binary hash of a visual search embedding for the query image with binary hashes of visual search embeddings for the images of the set of trusted digital images.

15 . The system of claim 9 , wherein the one or more processors are further configured to cause the system to generate one or more visual indicators from the fused feature vector utilizing one or more additional neural network layers, the one or more visual indicators identifying locations of editorial modifications in the query image relative to the trusted near-duplicate image.

16 . The system of claim 15 , wherein the one or more processors are configured to cause the system to generate the one or more visual indicators by:
generating a heat map from the fused feature vector utilizing a multilayer perceptron; and overlaying the one or more visual indicators on the query image based on the heat map.

17 . The system of claim 9 , wherein the one or more processors are configured to cause the system to align the query image with the trusted near-duplicate image by:
generating an optical flow between the query image and the trusted near-duplicate image utilizing an optical flow estimator; and warping the query image utilizing a de-warping unit based on the optical flow to generate the aligned query image.

18 . A computer-implemented method comprising:
aligning a first image with a second image to generate an aligned first image based on an optical flow between the first image and the second image; generating a first set of deep features for the aligned first image and a second set of deep features for the second image utilizing a neural network feature extractor; generating a fused feature vector by combining the first set of deep features and second set of deep features; generating a classification for modifications of the first image relative to the second image as benign, editorial, or a different image; and generating one or more visual indicators from the fused feature vector utilizing one or more neural network layers, the one or more visual indicators identifying locations of editorial modifications in the first image relative to the second image.

19 . The computer-implemented method of claim 18 , wherein generating the one or more visual indicators comprises:
generating a heat map from the fused feature vector utilizing a multilayer perceptron; and overlaying the one or more visual indicators on the first image based on the heat map.

20 . The computer-implemented method

Assignees

• Adobe Inc
• Univ Surrey

Inventors

• Collomosse John
• Black Alexander
• Bui Van Tu
• Jin Hailin
• Swaminathan Viswanathan

Classifications

• G06T3/18

  Image warping, e.g. rearranging pixels individually · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06T2207/20221

  Image fusion; Image merging · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20081

  Training; Learning · CPC title