Self-Supervised Anomaly Localization Using Synthetic Image Defects

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Anomaly detection in computer vision faces challenges due to the scarcity of labeled or unlabeled anomalous data and the fine-grained nature of differences between normal and anomalous patterns, leading to ineffective training of machine learning models.

Innovation Solution

A self-supervised learning method using a model trainer that generates augmented images by cropping and duplicating portions of normal images to simulate anomalies, employing techniques like duplication, rotation, and color jitter, enabling robust training without extensive anomalous data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If supervised classification is used for anomaly detection, then model training requires large amounts of labeled data, but labeled anomalous data is difficult to obtain due to the nature of the problem

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidamount of labeled anomalous data
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The system performs self-supervised learning by automatically generating synthetic anomalous samples from normal samples through augmentation techniques. The model trains itself by detecting anomalies in these self-generated samples, eliminating the need for external labeled anomalous data while maintaining detection accuracy

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system pre-generates synthetic anomalous samples before actual anomaly detection is needed. By creating augmented normal samples that contain simulated anomalies in advance, the model is prepared with diverse training data that covers various anomaly patterns, improving its ability to detect real anomalies

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If traditional anomaly detection methods are used, then training can be performed with limited data, but the model fails to capture fine-grained differences between normal and anomalous patterns

Engineering Contradiction:
Improvedetection of fine-grained anomaliesVSAvoidamount of training data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system crops normal samples into multiple smaller patches, treating each patch as an independent training sample. This segmentation allows the model to focus on local fine-grained patterns and defects, improving detection precision while utilizing the same limited data more effectively

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system applies various augmentation transformations (rotation, scaling, color jitter, noise addition) to normal samples to create diverse synthetic training data. These parameter changes simulate different anomaly conditions and help the model learn robust fine-grained features without requiring additional real anomaly data

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4244826B1Self-supervised learning for anomaly detection and localization
Publication Date: 2026.03.04 GOOGLE LLC
  • EP4244826B1 patent drawingFigure 1
  • EP4244826B1 patent drawingFigure 2
  • EP4244826B1 patent drawingFigure 3

AI summary

A method (500) for training a machine learning model (150) includes obtaining a set of training samples (112). For each training sample in the set of training samples, during each of one or more training iterations, the method includes cropping the training sample to generate a first cropped image (140 A), cropping the training sample to generate a second cropped image (140B) that is different than the first cropped image, and duplicating a first portion (210) of the second cropped image. The method also includes overlaying the duplicated first portion of the second cropped image on a second portion (220) of the second cropped image to form an augmented second cropped image (140BA). The first portion is different than the second portion. The method also includes training the machine learning model with the first cropped image and the augmented second cropped image.