Clustering Reference Images for Non-Destructive Testing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer vision systems face inefficiencies in detecting features in images due to the increasing computational workload with more reference images, where the incremental return decreases, and struggles with capturing variations such as texture and out-of-plane rotations.
Innovation Solution
The system clusters training images based on image-alignment data to select representative reference images, using techniques like dimensionality reduction and spectral clustering to maximize content variations while minimizing the number of images, thereby improving defect detection robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If more reference images are used in computer vision systems, then detection accuracy may improve, but computational workload increases and efficiency decreases
Solution Approach 1:
The patent extracts and removes redundant reference images from the dataset through clustering analysis. By grouping similar images into clusters and selecting only representative images from each cluster, the system eliminates excessive duplicate information while preserving detection accuracy, thereby reducing computational workload.
Solution Approach 2:
The patent transforms the reference image dataset by changing its composition parameters - reducing the total number of images while maintaining diversity through cluster-based selection. This parameter change optimizes the balance between having enough varied reference images for accurate detection and having fewer images for computational efficiency.
2Reliability
If more reference images are used to capture variations, then detection robustness improves, but the incremental return decreases
Solution Approach 1:
The patent extracts only the essential variations needed for robust detection by identifying representative images from each cluster. This extraction process removes redundant images that provide diminishing returns, keeping only those images that contribute meaningfully to capturing texture and rotation variations.
Solution Approach 2:
Instead of using all available reference images, the patent applies partial action by selecting a subset of representative images from each cluster. This partial selection is sufficient to capture the necessary variations for robust detection without the excessive computational cost of using all images.
3Device complexity
If traditional clustering is used without image alignment, then processing is simpler, but alignment-related variations are not captured
Solution Approach 1:
The patent performs preliminary image alignment before clustering by generating image-alignment data that captures transformations between images. This preliminary action ensures that alignment-related variations are accounted for before the clustering process, improving the reliability of detecting true content variations.
Solution Approach 2:
The patent introduces image-alignment data as an intermediary between the raw images and the clustering process. This intermediary layer processes and normalizes alignment variations, allowing the clustering algorithm to focus on meaningful content differences rather than being confounded by positional or rotational variations.
Data Source
AI summary
Devices, systems, and methods obtain training images; generate image-alignment data based on the training images; cluster the training images based at least in part on the image-alignment data, thereby generating clusters of training images; and select one or more representative images from the training images based on the clusters of training images.