# US 12,597,232 B2 · System for learning new visual inspection tasks using a few-shot meta-learning method

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,597,232 B2
- Document ID: US12597232B2
- Status: Granted
- Kind: B2
- Application number: 18/210,221
- Filed: 2023-06-15
- Published: 2024-12-19
- Granted: 2026-04-07
- Earliest claimed: 2023-06-15
- Assignees: Hitachi, Ltd.
- Applicants: HITACHI, Ltd.
- Inventors: Lasitha Sandaruwan Vidyaratne; Xian Yeow Lee; Mahbubul Alam; Ahmed Farahat; Dipanjan Ghosh; Maria Teresa Gonzalez Diaz; Chetan Gupta
- Web page: https://patentgrep.com/patent/US12597232B2

## CPC

- G06N 3/045: Combinations of networks
- G06V 10/762: using clustering, e.g. of similar faces in social networks
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/096: Transfer learning
- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06V 10/761: {Proximity, similarity or dissimilarity measures}
- G06V 10/7715: {Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods}
- G06V 10/7753: {Incorporation of unlabelled data, e.g. multiple instance learning [MIL]}
- G06N 3/0499: Feedforward networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

Systems and methods described herein which can involve for a first input of a plurality of labeled images of a new domain task, processing the first plurality of labeled images through a plurality of backbone snapshots, each of the backbone snapshots representative of a model trained across a plurality of other domain tasks, each of the plurality of backbone snapshots configured to output a first plurality of features responsive to the input; processing a second input of second plurality of unlabeled images through the plurality of backbone snapshots to output a second plurality of features responsive to the second input; and generating a representative model for the new domain task from the clustering and transformation of the first plurality of features and as associated from the clustered and transformed second plurality of features.

## Claims

### Claim 1

A method, comprising:
for an input of a first plurality of labeled images of a new domain task, processing the first plurality of labeled images through a plurality of backbone snapshots, each of the backbone snapshots representative of a model trained across a plurality of other domain tasks, each of the plurality of backbone snapshots configured to output a first plurality of features responsive to the input;
processing a second input of second plurality of unlabeled images through the plurality of backbone snapshots to output a second plurality of features responsive to the second input;
executing a feature transform process configured to transform the first plurality of features and the second plurality of features into similarity values that relate the first and the second plurality of features together based on similarity;
executing a clustering process to cluster the first plurality of features and the second plurality of features based on the similarity values; and
generating a representative model for the new domain task from the clustering and transformation of the first plurality of features and as associated from the clustered and transformed second plurality of features, wherein the new domain task is visual inspection, wherein the labeled images involve defect types for an object or task for the visual inspection, wherein the representative model is configured to output defects detected in the visual inspection.

### Claim 2 (depends on 1)

The method of claim 1, wherein the plurality of backbone snapshots are meta-trained in combination with a multilayer perceptron (MLP).

### Claim 3 (depends on 2)

The method of claim 2, wherein the plurality of backbone snapshots are meta-trained based on a classification loss and a contrastive loss.

### Claim 4 (depends on 1)

The method of claim 1, wherein features produced from the plurality of backbone snapshots are representative of industrial tasks or objects.

### Claim 5 (depends on 1)

The method of claim 1, wherein the each of the plurality of the backbone snapshots are configured with squeeze and excitation layers and anti-aliasing filters to facilitate cross-domain visual inspection.

### Claim 6 (depends on 1)

The method of claim 1, wherein the new domain task is associated with a first plurality of classes for the task, wherein the plurality of labeled images comprises examples for each of the plurality of classes.

### Claim 7

A non-transitory computer readable medium, storing instructions for executing a process, the instructions comprising:
for an input of a first plurality of labeled images of a new domain task, processing the first plurality of labeled images through a plurality of backbone snapshots, each of the backbone snapshots representative of a model trained across a plurality of other domain tasks, each of the plurality of backbone snapshots configured to output a first plurality of features responsive to the input;
processing a second input of second plurality of unlabeled images through the plurality of backbone snapshots to output a second plurality of features responsive to the second input;
executing a feature transform process configured to transform the first plurality of features and the second plurality of features into similarity values that relate the first and the second plurality of features together based on similarity;
executing a clustering process to cluster the first plurality of features and the second plurality of features based on the similarity values; and
generating a representative model for the new domain task from the clustering and transformation of the first plurality of features and as associated from the clustered and transformed second plurality of features, wherein the new domain task is visual inspection, wherein the labeled images involve defect types for an object or task for the visual inspection, wherein the representative model is configured to output defects detected in the visual inspection.

### Claim 8 (depends on 7)

The non-transitory computer readable medium of claim 7, wherein the plurality of backbone snapshots are meta-trained in combination with a multilayer perceptron (MLP).

### Claim 9 (depends on 8)

The non-transitory computer readable medium of claim 8, wherein the plurality of backbone snapshots are meta-trained based on a classification loss and a contrastive loss.

### Claim 10 (depends on 7)

The non-transitory computer readable medium of claim 7, wherein features produced from the plurality of backbone snapshots are representative of industrial tasks or objects.

### Claim 11 (depends on 7)

The non-transitory computer readable medium of claim 7, wherein each of the plurality of the backbone snapshots are configured with squeeze and excitation layers and anti-aliasing filters to facilitate cross-domain visual inspection.

### Claim 12 (depends on 7)

The non-transitory computer readable medium of claim 7, wherein the new domain task is associated with a first plurality of classes for the task, wherein the plurality of labeled images comprises examples for each of the plurality of classes.

### Claim 13

An apparatus, comprising:
a processor, configured to:
for an input of a first plurality of labeled images of a new domain task, process the first plurality of labeled images through a plurality of backbone snapshots, each of the backbone snapshots representative of a model trained across a plurality of other domain tasks, each of the plurality of backbone snapshots configured to output a first plurality of features responsive to the input;
process a second input of second plurality of unlabeled images through the plurality of backbone snapshots to output a second plurality of features responsive to the second input;
execute a feature transform process configured to transform the first plurality of features and the second plurality of features into similarity values that relate the first and the second plurality of features together based on similarity;
execute a clustering process to cluster the first plurality of features and the second plurality of features based on the similarity values; and
generate a representative model for the new domain task from the clustering and transformation of the first plurality of features and as associated from the clustered and transformed second plurality of features, wherein the new domain task is visual inspection, wherein the labeled images involve defect types for an object or task for the visual inspection, wherein the representative model is configured to output defects detected in the visual inspection.