Defect Inspection Device Using Machine Learning Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional defect inspection systems face challenges in accurately determining the presence of defects due to noise in feature extraction images, which can lead to erroneous judgments and decreased productivity, as the inspection conditions often differ from the learning conditions.
Innovation Solution
A defect inspection device and method that utilizes an identification part trained through machine learning to generate feature extraction images, specifies defect areas based on judgment parameters, and calculates image scores to update parameters such that the difference between inside and outside scores becomes significant, thereby reducing noise and improving judgment accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an identification part that has executed learning in advance is used for defect inspection, then defect detection capability is improved, but noise from reflection and shade causes erroneous judgments and requires time-consuming parameter adjustment
Solution Approach 1:
The system performs preliminary action by automatically calculating optimal judgment parameters using image scores before actual defect inspection begins. The setting part computes image scores based on pixel color density in binary images and determines parameters that maximize the difference between inside and outside area scores, eliminating the need for time-consuming manual parameter adjustment while maintaining high defect detection accuracy
Solution Approach 2:
The system implements self-service by enabling the identification part to automatically optimize its own judgment parameters using the image scoring mechanism. The setting part automatically calculates parameters based on the statistical properties of the inspection images themselves, allowing the system to self-adjust without external intervention or manual tuning, thereby reducing parameter adjustment time while preserving measurement precision
2Measurement precision
If judgment parameters are manually adjusted to separate noise and defects, then judgment accuracy is improved, but productivity decreases due to time consumption
Solution Approach 1:
The system replaces the mechanical manual adjustment process with an automated computational mechanism. The setting part uses image scoring based on pixel color density statistics to automatically calculate optimal judgment parameters, substituting the manual mechanical adjustment process with an automated algorithmic system that achieves the same noise-separation function without time loss, thereby maintaining judgment accuracy while improving productivity
Solution Approach 2:
The system implements parameter changes by dynamically calculating optimal judgment parameters based on image-specific characteristics. The setting part computes parameters that maximize the difference between inside and outside area scores for each inspection case, allowing the system to adapt parameters automatically to the specific image being analyzed, achieving high judgment accuracy without manual intervention and thus maintaining productivity
3Reliability
If the identification part is trained with learning image data, then defect recognition capability is improved, but inspection conditions may not match learning conditions causing noise
Solution Approach 1:
The system addresses condition mismatch by automatically calculating optimal judgment parameters based on the actual inspection image being analyzed. The setting part computes image scores and determines parameters that maximize the difference between inside and outside area scores for each specific image, allowing the system to adapt to different inspection conditions dynamically while maintaining the defect recognition reliability established during learning training
Data Source
AI summary
An image generating part generating feature extraction images by applying an identification part, which has completed learning, that has executed learning in advance to extract features using learning image data to an inspection image, an inspection part specifying an area corresponding to a defect on the basis of judgment parameters for judging presence/absence of a defect in the inspection target object and a binary image generated on the basis of the feature extraction images, and a setting part calculating an image score based on a density of a color of pixels of a setting image using the setting image that is the binary image in which an area corresponding to the defect is specified and updating the judgment parameters such that a difference between an image score of the inside of the area and an image score of the outside of the area becomes relatively large are included.