Facial Expression Labeling for Japanese Individuals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial expression recognition techniques, based on Ekman's theory, struggle to accurately annotate and classify facial expressions for Japanese individuals, often mislabeling emotions like anger, fear, sadness, and disgust due to differences in facial deformation and intensity compared to Westerners.
Innovation Solution
A facial expression label assignment apparatus that calculates representative feature quantities for a specific group, such as Japanese individuals, using action units (AU) data from the Facial Action Coding System (FACS), allowing for accurate label assignment by comparing target feature quantities with pre-calculated representative features, thereby improving the accuracy of facial expression recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If Ekman's theory is used for facial expression annotation, then the annotation process is simple and fast, but the labeling accuracy is low for Japanese individuals due to cultural and racial differences in facial expression deformation
Solution Approach 1:
The patent applies local quality by creating group-specific feature quantities tailored to different racial or cultural groups (e.g., Japanese vs. Westerners). Instead of using a universal annotation standard, the system adapts the feature extraction and labeling process to local characteristics of each group, such as different facial deformation patterns, thereby improving labeling accuracy for specific populations while maintaining efficient automated processing.
Solution Approach 2:
The patent changes parameters by introducing group-specific feature quantities that modify the standard facial expression analysis parameters. By adjusting the feature extraction parameters according to the target group's characteristics (e.g., different action unit weights or deformation thresholds), the system achieves accurate labeling for diverse populations without sacrificing annotation efficiency.
2Ease of operation
If visual classification based on Ekman's theory is used, then the annotation method is easy to implement, but wrong labels are assigned due to inability to clearly discriminate facial expressions in Japanese individuals
Solution Approach 1:
The patent replaces the mechanical visual classification system with an automated computational system that uses machine learning and group-specific feature quantities. This substitution eliminates human subjectivity and cultural bias in visual assessment, providing reliable automated labeling that adapts to different racial groups' facial expression characteristics while maintaining ease of operation through algorithmic processing.
Solution Approach 2:
The patent introduces an intermediary layer of group-specific feature quantities and machine learning models between the raw facial image and the final emotion label. This intermediary process transforms the simple visual classification into a multi-stage automated system that accounts for cultural differences, thereby improving reliability without complicating the overall annotation workflow for users.
3Ease of manufacture
If standard facial expression recognition models are trained on general data, then the model construction is straightforward, but the model performance is poor for specific groups like Japanese individuals due to differences in facial deformation patterns
Solution Approach 1:
The patent segments the training data and model construction process by racial or cultural groups, creating group-specific feature quantities and training datasets. Instead of using a single universal model, the system divides the recognition task into group-specific sub-tasks, allowing the model to learn the unique facial deformation patterns of each group (e.g., Japanese vs. Westerners) while maintaining a standardized overall framework for model construction.
Solution Approach 2:
The patent introduces dynamics by making the model adaptable to different groups through group-specific feature quantities. The system dynamically adjusts the feature extraction and model parameters based on the target group's characteristics, enabling the same base model architecture to achieve high accuracy across diverse populations by adapting to their specific facial expression patterns.
Data Source
AI summary
A facial expression label is assigned to face image data of a person with high accuracy. A facial expression data set storage unit (110) stores a facial expression data set in which the facial expression label is assigned to the face images in which people belonging to various groups show various facial expressions. A facial expression sampling unit (11) acquires a face image in which a person belonging to the desired group shows a desired facial expression. A representative feature quantity calculation unit (12) determines a representative feature quantity for each facial expression label from the face image of the desired group. The target data extraction unit (13) extracts target data from a facial expression data set. A target feature quantity calculation unit (14) calculates a target feature quantity from the target data. A corrected label determination unit (15) determines a corrected label to be assigned to the target data by comparing the target feature quantity and each of the representative feature quantities. The label assignment unit (16) assigns a corrected label to the target data.