Marker-Aware Image Anomaly Detection Without Labeled Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image anomaly detection methods using convolutional neural networks (CNNs) face challenges in training due to the need for labeled datasets, making it difficult to effectively detect anomalies in unlabeled data.
Innovation Solution
A processor-implemented method generates image and marker features by removing markers from input and reference images using an AI model, determining similarities between feature maps and patch similarities, and detecting anomalies based on these comparisons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural network (CNN) is used for image anomaly detection, then detection capability is improved, but training difficulty increases due to requirement of labeled datasets
Solution Approach 1:
The system performs self-supervised learning by automatically generating pseudo-labels from the image data itself. The CNN model learns to detect anomalies by comparing input images with reference images, where the reference images serve as self-generated training labels. This eliminates the need for external labeled datasets, allowing the system to train autonomously on unlabeled data while maintaining high detection accuracy.
2Measurement precision
If labeled datasets are used for CNN training, then detection accuracy is improved, but data preparation complexity increases
Solution Approach 1:
The system generates its own training data by creating reference images from normal manufacturing process images. These reference images automatically serve as labeled training data, where the normal images define what constitutes acceptable quality. This self-generating approach eliminates manual annotation work and complex data preparation processes while providing sufficient training data for accurate anomaly detection.
Solution Approach 2:
The system performs preliminary image processing to generate reference images and feature maps before actual anomaly detection. By pre-processing normal images to create reference feature maps and storing them as training data, the system prepares all necessary labeled data in advance. This preliminary action simplifies subsequent detection operations, as the training data is already prepared and organized for direct comparison with new input images.
3Measurement precision
If both image and marker features are analyzed, then detection accuracy is improved, but processing complexity increases
Solution Approach 1:
The system segments the analysis into two distinct feature extraction processes: image feature extraction from the main image content and marker feature extraction from specific marker regions. By dividing the processing into separate feature maps for images and markers, the system can independently analyze each type of feature and then combine the results. This segmentation reduces overall processing complexity by allowing specialized handling of each feature type while maintaining comprehensive detection accuracy.
Data Source
AI summary
A processor-implemented method includes generating an image feature of an input image and a marker feature of a marker marked on the input image, determining a comparison result of the image feature by comparing the image feature of the input image with a reference image feature of one or more reference images, determining a comparison result of the marker feature by comparing the marker feature with a reference marker feature of a reference marker on the one or more reference images, and detecting whether an anomaly is in the input image based on the comparison result of the image feature and the comparison result of the marker feature.