Defect Classification Analyzer With Feature-Space Similarity Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image classification techniques struggle to accurately classify defects on object surfaces and fail to recognize similarity relationships between classifications.
Innovation Solution
A defect analyzer that includes a feature extraction model trained on learned data to reduce distances between similar images, determines representative points for classifications, and visualizes relations using heat maps, scatter plots, or dendrograms to analyze defect classifications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image classification techniques are used, then classification can be performed, but it is difficult to accurately classify defects on object surfaces and recognize similarity relationships between classifications
Solution Approach 1:
The patent transforms the classification problem from traditional category-based classification to a geometric distance-based approach in a feature space. By mapping defect images to points in a multi-dimensional space and using distance metrics to determine similarity, the system achieves both accurate classification and recognition of similarity relationships simultaneously.
Solution Approach 2:
The patent changes the parameter representation from discrete class labels to continuous distance metrics in feature space. By representing defect similarity as Euclidean distances between feature vectors rather than discrete categories, the system enables both precise classification and continuous similarity assessment.
2Productivity
If only classification results are presented, then defects are classified, but it is difficult to recognize similarity relationships between classifications
Solution Approach 1:
The patent provides feedback by visualizing the geometric relationships between classified defects. The system not only outputs classification results but also displays distance metrics and spatial relationships in feature space, allowing users to understand similarity relationships and potentially refine classifications through iterative feedback.
Solution Approach 2:
The patent introduces distance metrics and feature space representations as intermediaries between the classification process and the final results. These intermediaries preserve and convey similarity relationship information that would otherwise be lost in discrete classification outputs.
3Extent of automation
If deep learning is used for image classification, then classification can be performed, but it is difficult to classify defects with variety occurring on object surfaces
Solution Approach 1:
The patent creates a universal classification framework that handles diverse defect types through a unified distance-based approach. The feature extraction and distance calculation mechanism works universally across different defect types, eliminating the need for defect-specific classification rules while maintaining high adaptability.
Data Source
AI summary
A classification of a defect is analyzed based on an image obtained by imaging the defect. A defect analyzer includes a model storage storing a feature extraction model that is trained based on learned data in which information indicating a classification of each defect is added to an image of the defect, the image of the defect being obtained by imaging the defect that occurs on a surface of an object, and the feature extraction model being configured to extract, from an input image, the image feature that decreases a distance between similar images; a representative-point determination unit configured to determine a representative point for each classification, based on one or more image features that are extracted from respective images of defects by using the feature extraction model; and a relation visualization unit configured to output information indicating a relation between classifications, based on one or more distances that are each between representative points.