Label Candidate Generation for Image Recognition Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The process of assigning labels to large amounts of captured images for image recognition training data is time-consuming and labor-intensive for operators, as existing technologies do not adequately address the workload associated with data labeling.
Innovation Solution
A learning device and system that acquires captured images, identifies candidate objects using an identification model, and displays these candidates as label options on a display device, allowing users to select appropriate labels, thereby reducing the workload for operators.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If operators manually assign labels to large amounts of captured images, then labeling accuracy can be maintained, but the time and labor required increases significantly
Solution Approach 1:
The identification model performs preliminary object recognition and generates candidate labels before the operator views the images. This preliminary action pre-processes the labeling task by providing predicted labels and confidence scores, so operators only need to verify or correct rather than create labels from scratch, significantly reducing labeling time while maintaining accuracy
Solution Approach 2:
The identification model acts as an intermediary between the captured images and the final labels. Instead of operators directly assigning labels to raw images, the model provides intermediate candidate labels with confidence scores, which then guide the operator's final label assignment decision, reducing the cognitive load and time required
2Measurement precision
If operators manually assign labels to captured images, then label quality can be ensured, but the workload and complexity of the process increases
Solution Approach 1:
The system performs self-service by automatically generating candidate labels and confidence scores through the identification model. This automation handles the complex task of object recognition and label generation, reducing the operational complexity for human operators who only need to review and confirm the automatically generated labels
3Productivity
If an identification model is used to generate candidate labels, then labeling workload is reduced, but the system complexity increases
Solution Approach 1:
The manual mechanical process of label assignment is replaced with an automated identification model that uses machine learning algorithms. This substitution increases productivity by automating the label generation process, while the system complexity is managed by integrating the model into the existing workflow where it operates as a automated assistant rather than a fully autonomous system
Data Source
AI summary
In generation of training data for image recognition, in order to reduce a user's workload of assigning a label to a captured image, a learning device includes a processor for performing operations to generate training data and a display device, wherein the processor is configured to acquire a captured image from an image capturing device; acquire one or more candidate objects recognized based on an identification model for a recognition target included in the captured image; and display, on the display device, information on the candidate objects as respective label candidates for the captured image.