Facial Image Blending Using Distance-Selected Training Pairs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for facial expression recognition face challenges in scaling training data due to high costs of annotation and the risk of overfitting, particularly when there are few classes of emotions and face images may contain components of multiple classes, leading to similar training data generation and poor generalization performance.
Innovation Solution
A method involving selecting face images with significant distance in their inference results to combine and generate new training data through alpha blending, ensuring diversity by using a second model trained on these combined images, which can include varying emotional elements and temporal changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training data is artificially generated using random alpha blending and mixup process, then the quantity of training data increases, but the diversity of training data decreases leading to poor generalization performance
Solution Approach 1:
The patent changes the selection parameter from random selection to distance-based selection using cosine similarity. By calculating the distance between inference results of candidate images and selecting those exceeding a threshold, the method ensures diverse training samples are chosen, improving generalization performance while maintaining data quantity increase.
Solution Approach 2:
The patent replaces the random mechanical selection process with a systematic distance-based selection mechanism. Instead of randomly pairing images, the system uses cosine similarity calculations to objectively measure and select diverse pairs, substituting randomness with a more sophisticated selection criterion.
2Ease of manufacture
If face images with similar characteristics are combined, then the training process becomes simpler, but the training data diversity decreases leading to overfitting
Solution Approach 1:
The patent introduces feedback through the distance calculation mechanism. By computing cosine similarity between inference results and using this metric to guide selection, the system creates a feedback loop that ensures diverse samples are chosen, preventing overfitting while maintaining training simplicity.
3Adaptability or versatility
If a large number of face images are used for training, then the model coverage increases, but the risk of overfitting increases when annotation costs are high
Solution Approach 1:
The patent performs preliminary action by pre-calculating inference results and their distances before the actual training process. By preparing diverse image pairs in advance using cosine similarity metrics, the system ensures high-quality diverse training data is ready, reducing overfitting risk while maintaining broad model coverage.
Data Source
AI summary
A storage unit stores therein first inference results respectively corresponding to a plurality of face images, which are output from a first model of outputting a first inference result in response to an input of a face image. A processing unit selects, from the plurality of face images, first and second face images whose distance calculated based their first inference results exceeds a threshold. The processing unit performs machine learning to train a second model of outputting a second inference result in response to an input of a face image, using a third face image obtained by combining the first and second face images.