Image Recognition Feature Database Management via Similarity Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face challenges in optimizing feature amount data for improved recognition accuracy due to limitations in storage capacity and the need for robustness against varying imaging conditions, with existing methods lacking clear methods for automatically determining and updating feature data for registration.
Innovation Solution
An image recognition apparatus and method that automatically determines the registration, disposal, and replacement of feature amount data by evaluating similarity using predetermined conditions, including a first condition where new data is not added if similar to existing data, and a second condition where new data is added if the number of registered data reaches an upper limit, ensuring optimal variation and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple pieces of feature amount data are registered for the same object to increase robustness against imaging condition variations, then recognition accuracy is improved, but database storage capacity is consumed
Solution Approach 1:
The patent changes the parameter of data selection by introducing similarity comparison metrics and registration conditions. Instead of registering all extracted feature data, the system evaluates each new feature data against existing registered data using similarity metrics, and only registers data that meets predetermined conditions (e.g., similarity threshold). This parameter-based filtering resolves the contradiction by selectively increasing data diversity while controlling storage consumption.
Solution Approach 2:
The patent extracts only the essential and useful feature amount data for registration by comparing new data with existing registered data. The similarity calculation unit identifies and extracts only those feature data that provide meaningful variation and contribute to robustness, discarding redundant data that would consume storage capacity without improving recognition accuracy. This selective extraction resolves the storage capacity vs. accuracy contradiction.
2Reliability
If the amount of registered feature amount data is increased to improve recognition accuracy, then robustness against imaging condition variations is improved, but the complexity of data management increases
Solution Approach 1:
The patent implements self-service data management through automated similarity comparison and registration determination. The system automatically evaluates new feature data against existing registered data, calculates similarity metrics, and determines registration necessity without manual intervention. This automated self-service mechanism resolves the data management complexity issue while maintaining improved recognition accuracy through systematic data selection.
Solution Approach 2:
The patent introduces feedback mechanisms where the registration determination unit continuously monitors the state of registered data and adjusts registration decisions based on similarity comparisons. The system provides feedback loops that evaluate whether new feature data should be registered by comparing it with existing data, ensuring optimal data selection. This feedback-based automation reduces management complexity while maintaining high recognition accuracy.
3Quantity of substance
If feature amount data with high similarity to existing data is registered, then the amount of registered data increases, but variation in registered feature amount data does not increase
Solution Approach 1:
The patent inverts the conventional approach by not registering data based on quantity alone, but rather by explicitly excluding data that does not meet variation requirements. The registration determination unit compares new feature data with existing registered data and determines non-registration when similarity exceeds predetermined thresholds. This inverted logic ensures that only data contributing to variation is registered, resolving the contradiction between data quantity and data variation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image recognition apparatus is provided, comprising: an extraction unit extracting feature amount data of a subject from an image; a database registering a plurality of pieces of feature amount data extracted from different images of one registered object; and a comparing unit identifying whether or not the subject is the registered object by comparing the feature amount data extracted by the extraction unit and the feature amount data of the registered object registered in the database, a registration unit, using an image for registration, adding feature amount data of the registered object to the database in accordance with a predetermined condition which includes a first condition: if new data, which is the feature amount data extracted from the image for registration, is similar to registered data, which is the feature amount data of the registered object already registered in the database, the new data is not added.