ReConPatch Feature Learning for Sparse Image Anomaly Patterns

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing anomaly detection methods face inefficiencies when dealing with high-dimensional image data, particularly in scenarios where abnormal patterns are rare, diverse, or labeled data is limited, leading to low accuracy and high computational costs, especially in identifying local patterns of varying sizes and shapes.

Innovation Solution

A patch feature learning method that involves acquiring a feature map from an image dataset, extracting patch features, performing feature representation learning using neural networks, and reconstructing ReConPatch features to enhance anomaly detection accuracy and efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If all data for the entire image is used to detect the anomaly, then the detection coverage is improved, but the data processing and computation costs increase

Engineering Contradiction:
Improveanomaly detection coverageVSAvoiddata processing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent divides the entire image into multiple patches and processes them independently. The feature extraction and anomaly detection are performed on individual patches rather than the whole image, reducing computational complexity while maintaining detection coverage across the entire image through aggregation of patch-level results.

Inventive Principle:
Principle #1Segmentation

2Reliability

If all data for the entire image is used to detect the anomaly, then the detection coverage is improved, but the computation costs increase

Engineering Contradiction:
Improveanomaly detection coverageVSAvoidcomputation cost
Core Design Contradiction:
ReliabilityVSPower

Solution Approach 1:

The patent segments the image into patches and performs feature extraction and anomaly detection on each patch separately. This segmentation reduces the computational burden by avoiding processing of the entire high-dimensional image data at once, while still achieving comprehensive detection coverage through combination of patch-level results.

Inventive Principle:
Principle #1Segmentation

3Ease of operation

If traditional methods are used to identify local patterns, then the processing simplicity is maintained, but the ability to identify local patterns of various sizes and shapes deteriorates

Engineering Contradiction:
Improveprocessing simplicityVSAvoidlocal pattern identification accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent divides the image into multiple patches, enabling the model to focus on local patterns within each patch. This segmentation approach allows the anomaly detection system to identify local patterns of various sizes and shapes by processing individual patches, improving local pattern recognition capability while maintaining relative processing simplicity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by performing feature extraction and anomaly detection specifically on local patches rather than treating the entire image uniformly. This allows the system to adapt to different local patterns and anomalies at various scales and positions within the image, improving the ability to detect diverse local anomalies.

Inventive Principle:
Principle #3Local quality

4Ease of manufacture

If labeled data is limited for training, then the data collection simplicity is maintained, but the model learning performance deteriorates

Engineering Contradiction:
Improvedata collection simplicityVSAvoidmodel learning performance
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent employs a two-stage training approach where the model is first pre-trained on large amounts of unlabeled data using self-supervised learning, and then fine-tuned on a smaller set of labeled anomaly data. This preliminary action on unlabeled data allows the model to learn general features and patterns before specializing in anomaly detection, improving learning performance even with limited labeled data.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements self-supervised learning where the model learns from unlabeled data by creating its own supervision signals through data augmentation and reconstruction tasks. This self-service learning mechanism enables the model to develop robust feature representations without requiring extensive manual labeling, thereby maintaining data collection simplicity while improving model performance.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4664360A1Patch feature learning method for anomaly detection, and system therefor
Publication Date: 2025.12.17 LG MANAGEMENT DEV INST CO LTD
  • EP4664360A1 patent drawingFigure 1
  • EP4664360A1 patent drawingFigure 2
  • EP4664360A1 patent drawingFigure 3

AI summary

A patch feature learning method and a patch feature learning system for anomaly detection according to an embodiment of the present disclosure perform patch feature-based learning on a predetermined pretrained model, based on an image data set for an anomaly detection target.