Self-Supervised Generative Model for Aerial Image Anomaly Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing anomaly detection systems for aerial inspections are inefficient and costly, particularly when detecting anomalies in new assets without historical imagery, and struggle to identify unique or novel anomalies due to their reliance on pre-defined features and statistical outlier analysis, which is not applicable to image data.

Innovation Solution

A self-supervised anomaly detection system using a generative model that reconstructs a digital image to its anomaly-free state, allowing for anomaly detection without prior images, by training a neural network to encode and decode images to eliminate anomaly features, thereby identifying deviations from the normal state.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If image change-based approach is used to compare current image to previous image, then anomaly detection can be performed, but it requires historical imagery which is not available for new assets and increases storage requirements

Engineering Contradiction:

Improveanomaly detection capabilityVSAvoidapplicability to new assets

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary training on anomaly-free images to build a generative model that learns the normal state of assets. This preliminary action enables the system to detect anomalies in new assets without requiring historical imagery, as the model has already encoded the expected normal appearance during the training phase

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The generative model creates a reconstructed copy of the input image representing the expected anomaly-free state. By comparing the original image with this synthesized copy, the system can identify anomalies without needing actual historical images, effectively using a generated replica instead of physical historical data

Inventive Principle:

Principle #26Copying

2Reliability

If feature-based classification approach is used with predefined anomaly databases, then known anomalies can be detected, but new types of anomalies or anomalies with different visual features may pass undetected

Engineering Contradiction:

Improvedetection of known anomaliesVSAvoiddetection of novel anomalies

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

Instead of training the model to recognize specific anomaly features as in traditional classification approaches, the system inverts the approach by training it to recognize only the normal anomaly-free state. The generative model learns what assets should look like, and any deviation from this learned normal state is automatically flagged as an anomaly, enabling detection of both known and novel anomaly types without requiring predefined anomaly databases

Inventive Principle:

Principle #13The other way round (Inversion)

3Adaptability or versatility

If statistical outlier detection is used for image data, then anomaly detection may be attempted, but image data has many more dimensions than access or financial data making it difficult to define outliers

Engineering Contradiction:

Improveapplication of outlier detection to imagesVSAvoiddimensionality of image data

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system transforms the high-dimensional image data into a lower-dimensional latent representation space through the generative model's encoding process. By learning the normal distribution in this compressed latent space, the system can effectively apply statistical outlier detection principles to image data despite its original high dimensionality, making anomaly detection feasible and computationally efficient

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

4Reliability

If conventional neural networks are trained to identify particular anomalies, then specific anomaly types can be detected, but the system requires additional information about what to output and cannot operate without supervision

Engineering Contradiction:

Improveanomaly identification accuracyVSAvoidtraining data requirements

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The generative model is self-supervised, meaning it automatically learns from the input images without requiring external supervision or additional information about anomaly types. The model encodes the input image and attempts to reconstruct it, learning the normal state distribution inherently from the data itself. This self-service capability eliminates the need for manually labeled training data or predefined anomaly categories, reducing complexity while maintaining detection effectiveness

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS10846531B2Image-based anomaly detection for aerial inspection using self-supervised learning

Publication Date: 2020.11.24 THE BOEING CO

AI summary

A method of automatically detecting anomaly from aerial images of an object of interest is provided. The method may include generating a data coding model corresponding to a category of assets by training a neural network with a training set of digital images depicting an asset in a state that is free from anomalies. The method may further include receiving a target digital image depicting a target asset, and reconstructing the target digital image using the data coding model to generate a decoded target digital image associated with the state that is free from anomalies. The data coding model may be self-supervised to learn to reconstruct itself to an anomaly-free state. The method may also include comparing the target digital image to the decoded target digital image to generate a difference map and, in response to a determination that the difference map depicts any anomaly, generating anomaly alert data.