Reconstruction-Based Anomaly Evaluation Without Defect Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection systems struggle to reliably identify defects or anomalies in scenes, particularly when using neural networks, due to difficulties in creating training datasets and consistently detecting anomalies in camera images.
Innovation Solution
A method involving calculating a reconstruction and a reference from a base datum, generating a discrepancy map, and determining a global anomaly score based on this map, with the reconstruction resembling the reference in the absence of anomalies, and triggering an alert when the score exceeds a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional neural network-based anomaly detection systems are used, then anomaly detection capability is provided, but training dataset creation becomes difficult and detection reliability decreases
Solution Approach 1:
Instead of training the neural network to directly detect anomalies, the patent inverts the approach by training it to reconstruct normal (defect-free) images. The anomaly detection is then performed by measuring the reconstruction error - when the error exceeds a threshold, an anomaly is detected. This inversion eliminates the need for annotated anomaly training data while maintaining detection capability.
Solution Approach 2:
The patent introduces a reconstruction model as an intermediary between the input image and anomaly detection. This reconstruction model serves as a mediator that learns the characteristics of normal images and enables anomaly detection through reconstruction error measurement, rather than requiring direct anomaly classification.
2Measurement precision
If neural networks are used for anomaly detection, then detection capability is achieved, but consistent and reliable anomaly detection becomes challenging
Solution Approach 1:
The patent extracts and focuses solely on learning the characteristics of normal (defect-free) images, separating this from anomaly detection. By taking out the anomaly detection task from the training process and handling it through reconstruction error measurement, the system achieves consistent and reliable detection with reduced complexity.
3Measurement precision
If comprehensive training datasets with anomalies are created, then detection accuracy improves, but training complexity and data preparation time increase
Solution Approach 1:
The patent inverts the traditional approach by training only on normal images rather than requiring comprehensive datasets with annotated anomalies. This eliminates time-consuming data preparation while maintaining detection accuracy through reconstruction error measurement.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
An anomaly evaluation method for detecting anomalies, the method comprising: S20) based on a base datum (BD), calculating a reconstruction (Rec) and a reference (Ref), the base datum (BD) representing a scene including part or all of one or more objects, the reconstruction (Rec) being a datum resembling to the base datum (BD), the reconstruction (Rec) being calculated with a reconstruction module (RM), the reference (Ref) being extracted from the base datum (BD); S30) calculating a Discrepancy Map (DM) representative of difference(s) between the reconstruction (Rec) and the reference (Ref) ; S40) calculating a global anomaly score (GAS) based on the Discrepancy Map (DM); the reconstruction module (RM) being configured to calculate reconstructions (Rec) such that, in the absence of anomaly(ies), the reconstruction (Rec) calculated based on a base datum (BD) is substantially identical to the reference (Ref) extracted from the base datum (BD).