Image Recognition Support Apparatus for Automated Label Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face challenges in accurately preparing correct labels for wide area images, leading to reduced accuracy in classifiers when incorrect labels are used in learning data.
Innovation Solution
An image recognition support apparatus that includes an image input unit, a pseudo label generation unit, and a new label generation unit. The pseudo label generation unit recognizes images using multiple types of image recognition models and generates pseudo labels, while the new label generation unit generates new labels based on these pseudo labels, thereby improving label reliability without manual confirmation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual label correction is performed by comparing preset label information with reliability output, then label accuracy is improved, but time consumption increases significantly
Solution Approach 1:
The system performs automatic label correction by having the image recognition unit generate reliability information and compare it with preset label information, eliminating the need for manual verification. The correction unit automatically identifies and corrects incorrect labels based on the reliability comparison, allowing the system to self-correct without human intervention.
Solution Approach 2:
The reliability information acts as an intermediary between the image recognition unit and the preset label information. By introducing this intermediate verification layer, the system can automatically identify discrepancies between generated labels and expected labels, enabling automated correction without direct manual comparison.
2Loss of time
If automatic label correction is performed based on reliability output, then time consumption is reduced, but classification accuracy decreases when reliability accuracy is insufficient
Solution Approach 1:
The system uses feedback from the reliability information to guide the correction process. The correction unit continuously monitors the reliability output and adjusts label corrections based on the confidence levels, ensuring that only labels with sufficient reliability evidence are automatically corrected, thereby maintaining classification accuracy while reducing time consumption.
Solution Approach 2:
Instead of automatically correcting all labels, the system performs partial correction only on labels where the reliability information indicates sufficient confidence. This selective approach prevents erroneous corrections while still achieving time savings on the majority of labels that can be confidently corrected.
3Productivity
If learning data includes incorrect labels, then productivity in creating learning datasets is improved, but classifier accuracy is reduced
Solution Approach 1:
The system performs preliminary label correction before the learning data is used for training classifiers. By pre-processing the labels to identify and correct errors using reliability comparison, the system ensures that the learning dataset contains accurate labels, thereby maintaining classifier accuracy while allowing rapid creation of learning datasets.
Data Source
AI summary
The invention supports creation of models for recognizing attributes in an image with high accuracy. An image recognition support apparatus includes an image input unit configured to acquire an image, a pseudo label generation unit configured to recognize the acquired image based on a plurality of types of image recognition models and output recognition information, and generate pseudo labels indicating attributes of the acquired image based on the output recognition information, and a new label generation unit configured to generate new labels based on the generated pseudo labels.