Printed Image Defect Discrimination With Unknown-Defect Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect detection systems in printed images can only identify defects that were part of the training data and fail to recognize unknown defects.
Innovation Solution
A defect discrimination apparatus and method using a machine learning model that includes a teacher image for known defects, a target image acquisition section, a discriminator for similarity analysis, and a learning section for additional or reinforcement learning when unknown defects are detected.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained using only known defect types, then the model can accurately identify trained defects, but it cannot detect unknown defects
Solution Approach 1:
The system incorporates a feedback mechanism where the learning section receives discrimination results from the discriminator and automatically performs additional learning when unknown defects are detected. This feedback loop enables the model to adapt to new defect types without manual retraining, resolving the contradiction between accurate identification of known defects and detection of unknown defects.
Solution Approach 2:
The learning model transitions from a static, fixed training state to a dynamic, self-updating system. The learning section continuously monitors discrimination results and performs additional learning when needed, allowing the model to adapt its knowledge base in real-time. This dynamic adjustment enables the model to maintain high accuracy for known defects while gaining capability to detect unknown defects.
2Adaptability or versatility
If the learning model is updated with new defect types, then it can detect more defect species, but the model complexity increases
Solution Approach 1:
The learning section performs self-service by automatically detecting when additional learning is needed and executing additional learning processes without external intervention. The system monitors its own performance through the discriminator and autonomously updates itself, reducing the need for complex manual training procedures while expanding defect detection capabilities.
Solution Approach 2:
The system prepares for future learning needs by continuously monitoring discrimination results and proactively performing additional learning when unknown defects are detected. This preliminary action approach ensures the model is always ready to handle new defect types, avoiding the need for complex post-hoc training procedures.
Data Source
AI summary
Prepare a learning model that has been trained to output similarity for each defect species by machine learning using a teacher image that is an image containing a defect occurring during printing and that is associated with a defect species in advance. Then, a target image is prepared for inspection by acquiring an image of printed matter that has been printed. By using the learning model with respect to this target image, the similarity of the defect present in the target image to a known defect species is acquired, and this similarity is used to discriminate the defect present in the target image as at least one of the known defect species. When updating the learning model based on this discrimination result, the learning model is made to perform machine learning for a defect species that is different from the discriminated defect species or that is associated with an unknown defect.