Feature Identification Device Using CNN Feature Value Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for identifying feature portions unique to each classification in image data require re-learning CNN parameters whenever a new classification is added, consuming computer resources and time.
Innovation Solution
A feature identification device that calculates similarity between reference and recognition feature values using a CNN, adjusting the recognition feature values based on this similarity, allowing for accurate identification without re-learning parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNN parameters are re-learned for each new classification, then identification accuracy is improved, but computational resources and time are consumed
Solution Approach 1:
The patent extracts and separates the feature extraction function from the classification function. The CNN model is used solely for extracting feature values from image data, while classification is performed by comparing these feature values against reference feature values stored in a database. This separation allows the feature extraction model to remain fixed while accommodating new classifications without re-learning.
Solution Approach 2:
The system segments the overall identification process into two independent stages: (1) feature extraction using CNN to generate feature values from input images, and (2) classification by comparing feature values with reference values in a database. This segmentation enables new classifications to be added by simply adding reference feature values to the database without modifying or re-training the CNN model.
2Measurement precision
If CNN parameters are re-learned for each new classification, then identification accuracy is improved, but computer resources are consumed
Solution Approach 1:
The patent extracts and separates the feature extraction function from the classification function. The CNN model is used solely for extracting feature values from image data, while classification is performed by comparing these feature values against reference feature values stored in a database. This separation allows the feature extraction model to remain fixed while accommodating new classifications without re-learning.
Solution Approach 2:
Instead of re-training the CNN model for each new classification, the system creates copies of reference images for each classification category and extracts reference feature values from these copies. These reference feature values are stored in a database, enabling new classifications to be added by simply adding reference data rather than re-learning the entire model.
Data Source
AI summary
A similarity calculation unit (213) treats each of a plurality of classifications as a target classification, and calculates a similarity between a reference feature value, which is a feature value extracted from image data of the target classification, and a recognition feature value, which is a feature value extracted from recognition target data, which is image data to be recognized. An influence calculation unit (214) calculates an influence on the similarity with regard to each partial image of the recognition target data by taking as input the similarity with regard to each of the classifications and the recognition feature value. A feature identification unit (215) changes the recognition feature value in accordance with the influence.