Computer vision image feature identification via multi-label few-shot model
AssigneeDST Technologies, Inc.
InventorsHui Peng Hu, Ramesh Sridharan
Life of this patent
Estimating the term…
Abstract
A technique making use of a few-shot model to determine graphical features present in an image based on a small set of examples with known graphical features. Where a support set including a number of images that each have a known combination of graphical features, the image recognition can identify unknown combinations of those graphical features in any number of query images. In an embodiment of the present disclosure examples of a filled-out form are used to interpret any number of additional filled out versions of the form.
Claims20
A method comprising:
- generating a binary matrix representing the presence of known graphic features in each of a set of support images;
- generating a query matrix representing a degree of similarity between graphic features of a query image and each of the set of support images;
- inverting the binary matrix into an inverted matrix;
- and
- identifying graphic features present in the query image based on a computation including the query matrix and the first inverted matrix.
A method of computer vision analysis of images that identifies graphical features in images via a few-shot model comprising:
- training a first model via a set of images, the set of images each including graphical features, wherein input of the first model is a few-shot;
- for each image of the set of images, generating a respective vector indicating the presence graphical features in each of the set of images, the first model including a first binary matrix having axes representing images of the set of images and known graphical features respectively, the first binary matrix indicating specific graphic features present in each of the set of images;
- inputting a query image to the trained machine learning model wherein the trained machine learning model outputs a query vector;
- generating a pairwise comparison matrix based on the query vector and the respective vectors, wherein the pairwise comparison matrix indicates a graphical features in common between the query image and each of the set of images respectively;
- and
- deriving the graphical features present in the query image based on an inverse of the first binary matrix and the pairwise comparison matrix.
A system of computer vision analysis of images that identifies graphical features in images via a few-shot model comprising:
- a graphic features model configured to generate a set of vectors describing graphical features in each of a set of images and a query image, where the graphical features present in the set of images are known graphical features to the graphical features model and the graphic features present in the query image are unknown to the graphics features model;
- a relation network model configured to generate a query matrix representing similarity of graphical features between the query image and each of the set of images respectively;
- a processor;
- and
- a memory including instructions configured to cause the processor to:
- invert a binary truth matrix into a first inverted matrix, the binary truth matrix including information describing known graphical features of each of the set of images;
- and
- derive graphic features present in the query image based on a computation including the binary query matrix and the first inverted matrix.
Dependencies are inferred from the claim text.