Granted Patent B2

US 10,289,909 · App. 15/450,620 · Granted May 14, 2019

Inventors: Fabien Baradel (Chambery, FR); Boris Chidlovskii (Meylan, FR); Gabriela Csurka (Crolles, FR)

Assignee: Xerox Corporation

G06K9/00624G06K9/4628G06K9/6262G06K9/6271

Abstract

Claims (25)

1. A method for classifying an image, comprising:

receiving, by a processor, one or more images associated with a source domain and one or more images associated with a target domain;

identifying, by the processor, a plurality of source layers, wherein each one of the plurality of source layers comprises one or more source domain features based on the one or more images associated with the source domain via a plurality of deep neural networks;

identifying, by the processor, a plurality of domain layers, wherein each one of the plurality of domain layers comprises one or more target domain features based on the one or more images associated with the target domain via the plurality of deep neural networks;

training, by the processor, a plurality of conditional maximum mean discrepancy (CMMD) engines based on a difference between the one or more source domain features and the one or more target domain features, wherein each one of the plurality of CMMD engines receives the one or more source domain features from a respective source layer of the plurality of source layers and the one or more target domain features from a respective domain layer of the plurality of domain layers, wherein a number of the plurality of CMMD engines is equal to a number of the plurality of deep neural networks;

applying, by the processor, the plurality of CMMD engines to the one or more images associated with the target domain to generate one or more labels for each unlabeled target image of the one or more images associated with the target domain; and

classifying, by the processor, each one of the one or more images in the target domain using the one or more labels.

2. The method of claim 1 , wherein the plurality of CMMD engines is optimized by a gradient descent algorithm.

3. The method of claim 1 , wherein the image comprises text.

4. A non-transitory computer-readable medium storing a plurality of instructions, which when executed by a processor, causes the processor to perform operations for classifying an image comprising:

receiving one or more images associated with a source domain and one or more images associated with a target domain;

identifying a plurality of source layers, wherein each one of the plurality of source layers comprises one or more source domain features based on the one or more images associated with the source domain via a plurality of deep neural networks;

identifying a plurality of domain layers, wherein each one of the plurality of domain layers comprises one or more target domain features based on the one or more images associated with the target domain via the plurality of deep neural networks;

training a plurality of conditional maximum mean discrepancy (CMMD) engines based on a difference between the one or more source domain features and the one or more target domain features, wherein each one of the plurality of CMMD engines receives the one or more source domain features from a respective source layer of the plurality of source layers and the one or more target domain features from a respective domain layer of the plurality of domain layers, wherein a number of the plurality of CMMD engines is equal to a number of the plurality of deep neural networks;

applying the plurality of CMMD engines to the one or more images associated with the target domain to generate one or more labels for each unlabeled target image of the one or more images associated with the target domain; and

classifying each one of the one or more images in the target domain using the one or more labels.

5. The non-transitory computer-readable medium of claim 4 , wherein the plurality of CMMD engines is optimized by a gradient descent algorithm.

6. The non-transitory computer-readable medium of claim 4 , wherein the image comprises text.

7. A method for classifying an image, comprising:

receiving, by a processor of an image classification system, one or more images associated with a source domain where the image classification system was previously deployed and trained and one or more images associated with a target domain where the image classification system is re-deployed;

obtaining, by the processor, one or more source domain features from the one or more images associated with the source domain based on the image classification system that was previously deployed and trained, wherein the one or more source domain features are identified in a plurality of source layers by a plurality of deep neural networks;

identifying, by the processor, a plurality of domain layers, wherein each one of the plurality of domain layers comprises one or more target domain features based on the one or more images associated with the target domain via the plurality of deep neural networks;

training, by the processor, a plurality of conditional maximum mean discrepancy (CMMD) engines based on a shift between the one or more source domain features and the one or more target domain features, wherein each one of the plurality of CMMD engines receives the one or more source domain features from a respective source layer of the plurality of source layers and the one or more target domain features from a respective domain layer of the plurality of domain layers, wherein a number of the plurality of CMMD engines is equal to a number of the plurality of deep neural networks;

applying, by the processor, the plurality of CMMD engines to the one or more images associated with the target domain to generate one or more labels for each unlabeled target image of the one or more images associated with the target domain; and

classifying, by the processor, each one of the one or more images in the target domain using the one or more labels.

Assignments (10)

Continuity (1)