Anomaly Detection Models Using Aligned Reference Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing nondestructive testing techniques struggle to efficiently detect anomalies in images without causing damage to the objects, particularly in quality-control processes where defects need to be identified accurately.
Innovation Solution
A device and method that involve obtaining training images, selecting reference images, generating aligned training images, and creating anomaly-detection models based on these alignments. The method further includes aligning test images to reference images, generating error maps, and combining them to produce a composite error map for anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple anomaly-detection models are generated based on different reference images, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent divides the anomaly detection task into multiple independent anomaly-detection models, each trained on a different reference image. This segmentation allows each model to specialize in detecting anomalies relative to its specific reference, improving overall detection precision while maintaining manageable complexity through modular model structures
Solution Approach 2:
The system creates multiple anomaly-detection models that can be applied to the same test image from different reference perspectives. Each model serves a specific function (comparing against one reference image), but collectively they provide universal anomaly detection capability across varying reference conditions
2Reliability
If training images are aligned to multiple reference images, then reliability is improved, but loss of time increases
Solution Approach 1:
The patent performs image alignment and anomaly-detection model generation during the training phase before actual anomaly detection is needed. By pre-aligning training images to multiple reference images and pre-training multiple models, the system establishes reliable detection capabilities in advance, so that during actual operation only inference is needed, reducing real-time processing time
Solution Approach 2:
The system creates multiple copies of anomaly-detection models, each trained on aligned images from different reference perspectives. These model copies can be deployed in parallel or selected based on the specific detection task, improving reliability through multiple verification paths while optimizing time by choosing appropriate models for different scenarios
Data Source
AI summary
Some devices, systems, and methods obtain training images; select a first reference image and a second reference image from the training images; generate a first set of aligned training images, wherein generating the first set of aligned training images includes aligning the training images to the first reference image; generate a first anomaly-detection model based on the first set of aligned training images; generate a second set of aligned training images, wherein generating the second set of aligned training images includes aligning the training images to the second reference image; and generate a second anomaly-detection model based on the second set of aligned training images.