Context-Based Defect Detection Using Image Patch Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect detection methods are limited by reliance on known defect libraries, requiring significant resource investment and not accounting for unknown defects.
Innovation Solution
A defect detection method based on self-similarities in image patches, evaluating changes in relationships between neighboring image patches without prior knowledge of defects, using similarity metrics like cosine similarity or patch-by-patch convolution to produce feature vector representations and determine defect status.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a defect library is used for defect detection, then known defects can be detected, but the detection capabilities are limited and significant resources are required to build and maintain the library
Solution Approach 1:
The system uses self-similarity metrics to automatically detect defects by comparing each image patch to all other patches in the image. The method computes similarity scores without requiring external defect libraries or training data, allowing the system to self-diagnose and detect unknown defect types through statistical analysis of patch relationships.
Solution Approach 2:
The invention changes the detection parameter from pattern matching (requiring defect libraries) to similarity score distribution analysis. By calculating z-scores based on the distribution of similarity scores across all patches, the system adapts to any defect type dynamically without predefined categories, transforming the detection approach from static library-based to dynamic statistic-based.
2Measurement precision
If a defect library is built and maintained, then known defects can be detected accurately, but significant resources are consumed for building and updating the library
Solution Approach 1:
The system performs self-diagnosis by computing similarity scores between all image patches and analyzing the statistical distribution of these scores. This self-service approach eliminates the need for manual creation and maintenance of defect libraries, reducing resource investment while maintaining high detection accuracy through automatic anomaly detection.
Solution Approach 2:
Instead of copying and storing defect patterns in a library, the system creates a statistical model of normal patch relationships by analyzing similarity score distributions. This probabilistic copying approach captures the essence of defect detection without requiring explicit defect examples, reducing resource requirements while maintaining detection precision.
3Adaptability or versatility
If known defect patterns are used for detection, then specific defects can be identified, but unknown defects cannot be detected
Solution Approach 1:
The system achieves universality by using general similarity metrics that apply to all patch comparisons regardless of defect type. The self-similarity approach creates a single detection mechanism that works for known and unknown defects alike, eliminating the need for multiple specialized detectors or defect-specific algorithms.
Solution Approach 2:
Instead of asking 'does this patch match a known defect pattern?', the system inverts the question to 'does this patch deviate from the normal distribution of patch similarities?' This inversion allows the system to detect unknown defects by identifying statistical anomalies rather than matching against predefined patterns, thereby expanding detection coverage to all possible defect types.
Data Source
AI summary
A method for context based detection of defects of manufactured items, the method may include (i) receiving an image of an evaluated manufactured item (EMI), the EMI was manufactured by a manufacturing process; (ii) generating EMI patches representations that are related to the EMI; wherein the EMI patches representations of the EMI are selected out of (a) representations of patches of the image of the EMI, or (b) patches of a representation of the image of the EMI; (iii) calculating EMI patches representations scores, wherein an EMI patch representation score of a certain EMI patch representation is determined based on similarities between the certain EMI patch representation and other EMI patch representations; and (iv) determining a defect related status of the EMI based on at least some of the EMI patches representations scores and on at least one similarity related values