Anomaly Detection Using Image Augmentation and Classifier Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in detecting anomalies in high-yield manufacturing processes where securing sufficient anomaly data is difficult due to a low defect rate.
Innovation Solution
A processor-implemented method with image analysis that involves receiving a test image, generating augmented images, determining classification prediction values using a classifier, calculating a detection score based on these values, and determining whether the test image corresponds to anomaly data based on the detection score and a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If anomaly detection is performed using deep-learning-based algorithms requiring both normal and anomaly data, then detection accuracy can be improved, but it becomes difficult to secure sufficient anomaly data due to low defect rate in high yield manufacturing processes
Solution Approach 1:
The patent inverts the traditional anomaly detection approach by training the classifier only on normal data and using it to detect anomalies through uncertainty measurement. Instead of requiring anomaly examples for training, the system leverages the classifier's confidence levels when presented with normal versus anomalous inputs, effectively solving the data scarcity problem while maintaining detection accuracy
Solution Approach 2:
The patent introduces an intermediary mechanism (uncertainty measurement based on classification prediction values) that bridges the gap between normal data training and anomaly detection. This intermediary allows the system to identify anomalies without direct anomaly examples by measuring how uncertain the classifier is when processing test images
2Reliability
If image augmentation is applied to generate diverse training samples, then the classifier's robustness can be improved, but the processing time and computational complexity increase
Solution Approach 1:
The patent applies image augmentation techniques (rotation, flipping, cropping, color jittering) during the training phase to pre-expose the classifier to varied conditions. This preliminary action builds robustness into the model before deployment, allowing faster and more reliable anomaly detection during actual operation without requiring real-time augmentation
3Measurement precision
If classification prediction values are used to determine anomaly detection scores, then detection accuracy can be improved, but the computational complexity of the system increases
Solution Approach 1:
The patent extracts only the essential information needed for anomaly detection from the classification output - specifically the prediction values and their entropy - rather than utilizing the full complexity of deep learning model internals. This extraction approach maintains high detection accuracy while significantly reducing computational complexity by focusing on the most informative metrics
Data Source
AI summary
A processor-implemented method with image analysis includes: receiving a test image; generating a plurality of augmented images by augmenting the test image; determining classification prediction values for the augmented images using a classifier; determining a detection score based on the classification prediction values; and determining whether the test image corresponds to anomaly data based on the detection score and a threshold.