Anomaly Map Fusion for Image Defect Detection With Sparse Defect Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection systems face challenges in accurately identifying anomalies in images, particularly in environments with low availability of defective images, such as in semiconductor manufacturing, where training with diverse and small numbers of defects is difficult.
Innovation Solution
A method involving a generative AI model to generate a normal image similar to an input image, followed by generating pixel-level and structural feature anomaly maps, merging these maps to create a multi-anomaly map, and using this map to detect anomalies, which includes preprocessing to reduce noise and enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional anomaly detection systems use distance-based methods with limited defective images, then the system can operate with minimal training data, but the detection accuracy deteriorates due to inability to learn diverse defect characteristics
Solution Approach 1:
Instead of directly learning from defective images (traditional approach), the patent inverts the approach by learning from normal images and generating synthetic defective images through anomaly addition. This allows the system to overcome the limitation of scarce defective images by creating diverse training data from abundant normal images, thereby improving detection accuracy without requiring large quantities of actual defective samples
Solution Approach 2:
The patent creates synthetic copies of defective images by adding anomalies to normal images. These generated defective images serve as training data, enabling the system to learn defect characteristics without relying on scarce real defective images. The copying process generates diverse defect variations that improve the model's ability to detect various anomaly types
2Device complexity
If the system focuses only on pixel-level difference detection, then the detection process is simple and fast, but the detection precision deteriorates due to noise and lack of structural context
Solution Approach 1:
The patent segments the anomaly detection process into two distinct components: pixel-level difference detection (capturing fine-grained anomaly details) and structural feature detection (capturing contextual patterns). By dividing the detection task into these separate analyses, the system achieves comprehensive precision while managing complexity through modular processing of different feature types
Solution Approach 2:
The patent merges pixel-level anomaly maps with structural feature anomaly maps to create a comprehensive detection result. This combination integrates the detailed local information from pixel comparison with the contextual information from structural analysis, thereby improving overall detection precision by leveraging the strengths of both approaches
3Productivity
If the system processes images without preprocessing, then the processing speed is fast, but the reliability deteriorates due to noise interference in the anomaly maps
Solution Approach 1:
The patent applies preprocessing operations (such as filtering and normalization) to images before anomaly detection. This preliminary action removes noise and enhances image quality in advance, ensuring that subsequent anomaly detection operates on cleaned data. The preprocessing step improves reliability by eliminating interference factors before the main detection process begins
Data Source
AI summary
A method and an apparatus are provided to detect an anomaly in an image through generating a first anomaly map related to pixel-level features of the input image and a second anomaly map related to structural features of the input image based on the input image and a non-defect image similar to the input image, and merging the first anomaly map and the second anomaly map to generate a multi-anomaly map.