Normal Vector Selection for Memory-Efficient Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional anomaly detection methods in manufacturing require large memory and processing resources due to the need for storing and calculating distances between inspection vectors and normal vectors, which can be inefficient and reduce recognition accuracy when reducing the number of normal images.
Innovation Solution
A method that generates feature vectors for each position of a normal image using a neural network, selects acquisition targets based on image processing, and stores only relevant feature vectors as normal vectors, reducing unnecessary memory and processing by eliminating unneeded feature vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of normal images are collected and stored to maintain recognition accuracy, then recognition accuracy is improved, but memory requirements and processing amount increase significantly
Solution Approach 1:
The patent extracts only the essential feature vectors from normal images that are most relevant for anomaly detection, rather than storing all normal images or all feature vectors. This selective extraction reduces memory requirements while preserving the key information needed for accurate recognition.
Solution Approach 2:
The patent applies local quality by differentiating the importance of different feature vectors and positions in the image. Instead of treating all feature vectors equally, it identifies and retains only those with higher importance for anomaly detection, optimizing the balance between memory usage and detection accuracy.
2Measurement precision
If all normal feature vectors are stored and distance calculations are performed for each inspection vector, then detection accuracy is maintained, but processing amount increases
Solution Approach 1:
The patent extracts and retains only the most important normal feature vectors for comparison during inspection. By removing redundant or less important feature vectors from the comparison set, it significantly reduces the number of distance calculations required while maintaining effective anomaly detection.
Solution Approach 2:
The patent applies partial action by performing distance calculations only with a selected subset of normal feature vectors rather than all available vectors. This partial comparison approach reduces processing load while still achieving sufficient detection accuracy for practical applications.
3Quantity of substance
If the number of normal vectors is reduced to decrease memory and processing requirements, then memory and processing amounts are reduced, but recognition accuracy may deteriorate
Solution Approach 1:
The patent ensures that reduced sets of normal vectors maintain high recognition accuracy by selectively retaining vectors with higher importance or representativeness. This quality-based selection ensures that fewer vectors can still provide accurate anomaly detection performance.
Solution Approach 2:
The patent changes the parameters of the normal vector set by selecting and retaining only those vectors that meet certain criteria for importance or representativeness. This parameter-based filtering reduces the number of vectors while preserving the essential characteristics needed for accurate detection.
Data Source
AI summary
According to one embodiment, a normal vector set creation apparatus generates a feature vector for each position of a normal image by performing feature amount extraction processing using a neural network on the normal image, generates position information indicating whether or not a feature vector at each position of the normal image is an acquisition target by performing image processing on the normal image, selects a feature vector of the acquisition target from among the generated feature vectors based on the position information, and stores the selected feature vector of the acquisition target as a normal vector.