Image Recognition Model Selection via Self-Generated Recall Results
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for selecting an optimum learning model for contour line extraction in semiconductor pattern inspection require a correct answer value or degree of certainty, which is time-consuming and not applicable when multiple types of learning models are involved.
Innovation Solution
An image recognition device and method that includes a feature extraction learning model group, a recall learning model group, a feature amount extraction unit, a data-to-data recall unit, and a learning model selection unit, which selects the feature extraction learning model based on minimizing the difference between the feature amount and the recall result, without requiring a correct answer value or degree of certainty.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the method of PTL 1 is used to select an optimum learning model by prediction error, then the selection accuracy is improved, but the preparation time and man-hours increase significantly due to requiring correct answer values for each image
Solution Approach 1:
The system uses the learning models themselves to generate recall results that serve as the basis for selection, eliminating the need for external correct answer value assignment. Each learning model generates its own recall result, and the selection is made based on comparing these self-generated results, thus the system serves itself without requiring manual preparation of reference data.
Solution Approach 2:
The recall result acts as an intermediary between the learning model output and the selection criterion. Instead of directly comparing learning model outputs with manually assigned correct answers, the system introduces recall results as an intermediate representation that can be automatically generated and compared, facilitating automated model selection without manual intervention.
2Extent of automation
If the method of PTL 2 is used to select an optimum learning model by degree of certainty, then the automation is improved, but the applicability decreases when multiple types of learning models are involved due to different certainty definitions
Solution Approach 1:
The recall result provides a universal selection criterion that works across different types of learning models. Instead of using model-specific metrics like degree of certainty that have different meanings for different model types, the system uses the recall result—a universal representation that can be generated by any learning model—to enable consistent automated selection across diverse model types.
3Manufacturing precision
If multiple learning models are prepared for different image types, then the contour extraction performance is improved, but the complexity of model selection increases
Solution Approach 1:
The system uses feedback from the recall results to automatically determine the optimal learning model. Each learning model generates a recall result that serves as feedback about its performance characteristics, and this feedback is used to automatically select the best model without complex manual evaluation, thus managing the complexity of having multiple models while maintaining high extraction performance.
Data Source
AI summary
In order to select an optimal learning model for an image when inference is carried out in the extraction of a profile line using machine learning, without requiring a correct value or degree of certainty, a feature extraction learning model group containing a plurality of learning models is used for feature extraction. A recall learning model group containing recall learning models is paired with the feature extraction learning models. A feature amount extraction unit for referencing a feature extraction learning model and extracting a feature amount from input data; a data-to-data recall unit for referencing a recall learning model and outputting a recall result with the feature amount subjected to dimensional compression; and a learning model selection unit for selecting a feature extraction learning model from the feature extraction learning model group under the condition that the difference between the feature amount and the recall result is minimized are provided.