Iterative Partial Image Selection for Training Model Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image prediction techniques face challenges when labels are given only for entire images, as it is unclear which parts of the image to use for training, leading to reduced prediction accuracy.
Innovation Solution
A training device and method that generates partial images from the input image, selects training partial images based on predicted values, and trains a prediction model to predict the presence of a feature in these partial images, iteratively updating the training data to improve model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If prediction is performed in units of partial image generated by dividing image into plurality of portions, then prediction coverage is improved, but training accuracy deteriorates when labels are given only for entire images
Solution Approach 1:
The patent segments the image into multiple partial images (e.g., grid division into 4 or more regions) to enable localized prediction. This segmentation allows the model to focus on specific regions while maintaining overall prediction coverage, resolving the contradiction between comprehensive coverage and localized training accuracy.
Solution Approach 2:
The patent applies partial action by selecting only specific partial images for training based on their predicted values, rather than using all partial images. This selective approach ensures that training focuses on the most informative regions, improving training accuracy while maintaining the benefits of comprehensive segmentation for prediction coverage.
2Productivity
If training is performed using only the partial image of the highest predicted value, then training efficiency is improved, but prediction accuracy deteriorates
Solution Approach 1:
The patent selects a plurality of partial images for training based on their predicted values, rather than using only the single highest predicted value. This partial action approach maintains training efficiency by focusing on the most informative regions while improving prediction accuracy by incorporating multiple high-value regions into training data.
Solution Approach 2:
The patent implements feedback by using the prediction results from partial images to guide the selection of training data. The predicted values from initial predictions are fed back to determine which partial images should be used for training, creating an iterative process that improves both efficiency and accuracy.
3Ease of manufacture
If labels are given only for entire images, then data preparation simplicity is improved, but training reliability deteriorates due to uncertainty about which image parts to use
Solution Approach 1:
The patent segments the labeled entire image into multiple partial images, allowing the model to learn from spatial relationships within the image. This segmentation transforms the training process to be less dependent on precise pixel-level annotations while maintaining reliability through the structured division of the labeled data.
Solution Approach 2:
The patent applies self-service by using the model's own predictions to automatically select training data. The system generates predictions for partial images and uses these predictions to determine which regions should be used for training, eliminating the need for manual annotation of specific regions while maintaining training reliability.
Data Source
AI summary
The partial image generation means generates partial images from the input image, and the partial image selection means selects a plurality of training partial images to be used as training data from all the generated partial images. The training means trains a prediction model for predicting a probability that a predetermined feature is included in the selected training partial images. The prediction means performs prediction for all the partial images using the trained prediction model. Then, the partial image selection means selects the plurality of training partial images to be used in a next training based on predicted values for all the partial images. The training of the prediction model by the training means is repeated while the training partial images are updated.