Training Data Generation Using Annotated Processed and Raw Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based image recognition technologies face challenges in utilizing training data that is optimized for human visibility rather than machine learning, leading to suboptimal performance in visual recognition tasks.
Innovation Solution
A system that generates training data by reflecting annotation results from images optimized for human visibility onto raw images, creating a combination of processed and unprocessed image data for machine learning, and records this data with metadata for multimodal learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If annotation results from processed images are used for training, then ease of operation is improved, but manufacturing precision deteriorates
Solution Approach 1:
The patent segments the training data into two distinct components: processed images (second image data) used for annotation and raw images (first image data) used for training. This segmentation allows annotators to work with enhanced images while the model trains on original images, resolving the contradiction between annotation ease and training accuracy.
Solution Approach 2:
The patent introduces processed image data as an intermediary medium. This intermediary serves as a bridge that makes annotation easier while the connection to raw image data ensures training precision. The processed images act as a mediator that facilitates the annotation process without compromising the quality of training data.
2Ease of operation
If only processed image data is used for training, then ease of operation is improved, but reliability deteriorates
Solution Approach 1:
The patent clearly separates the annotation source (processed images) from the training source (raw images). This segmentation ensures that the convenience of annotated processed images does not compromise the reliability of training, as the model learns from unprocessed original data that preserves all diagnostic information.
Solution Approach 2:
Processed images serve as an intermediary that facilitates accurate annotation while the raw images maintain reliability for training. The intermediary processed images enable annotators to achieve high accuracy, while the connection to raw images ensures the trained model maintains reliability for clinical inference.
3Loss of information
If processed and raw image data are combined, then information completeness is improved, but device complexity increases
Solution Approach 1:
The patent segments the data processing pipeline into distinct stages: annotation processing on second image data and training processing on first image data. This segmentation manages complexity by handling different data types in separate, well-defined processes rather than mixing them throughout the entire system.
Solution Approach 2:
The system uses a unified training data generation apparatus that handles both processed and raw image data through multi-functional processing units. The same apparatus performs annotation reflection, data pairing, and training set generation, reducing overall system complexity despite handling multiple data types.
Data Source
AI summary
A training data generation device includes a computer, and a computer-readable storage medium. The computer is configured to: receive an input of an annotation for second image data obtained by imaging an observation target; reflect a result of the annotation in first image data that is related to the same observation target as the observation target of the second image data, the first image data having a different at least one of imaging mode and display mode from the second image data; and generate training data for creating an inference model by using the first image data and the result of the annotation reflected in the first image data, the first image data including image data of a plurality of images, and the second image data being image data of an image obtained by combining the plurality of images included in the first image data.