Multi-dimensional Facial Beauty Prediction via Shared Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Facial beauty prediction systems face challenges due to imbalanced data samples, where classifiers skew towards majority class samples, leading to inefficiencies in classifying minority class samples and increased average classification costs.
Innovation Solution
A multi-dimensional task facial beauty prediction method that uses a shared feature extraction network and sub-task networks, where facial images are marked with multi-dimensional labels to supervise the networks, and parameters are optimized using a back propagation algorithm to address the imbalance, allowing for effective classification and prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single classifier is used for facial beauty prediction, then the device complexity is low, but the classification accuracy deteriorates due to data sample imbalance
Solution Approach 1:
The patent divides the single facial beauty classification task into multiple sub-tasks by creating several sub-classifiers, each responsible for predicting specific beauty dimensions (e.g., eye beauty, nose beauty, mouth beauty). This segmentation allows each sub-classifier to focus on specific features, improving overall classification accuracy while managing complexity through modular design.
Solution Approach 2:
The patent employs a shared feature extraction network that serves multiple sub-classifiers simultaneously. This multi-functional component extracts common facial features once and provides them to all sub-classifiers, reducing redundant computation and managing overall system complexity while maintaining high classification accuracy across multiple beauty dimensions.
2Reliability
If a classifier is trained on imbalanced data samples, then the training process is simple, but the classification effectiveness for minority class samples deteriorates
Solution Approach 1:
By segmenting the classification task into multiple sub-tasks with specialized sub-classifiers, the patent improves reliability for minority class samples. Each sub-classifier can be trained with balanced data for its specific dimension, overcoming the imbalance problem in the overall dataset while maintaining training feasibility through focused, dimension-specific learning.
Solution Approach 2:
The patent applies local quality by creating specialized sub-classifiers for different facial beauty dimensions, where each sub-classifier has optimized parameters and features tailored to its specific task. This local optimization ensures high reliability for each dimension's classification, including minority classes, while the overall training process remains manageable through modular, independent training of each sub-classifier.
3Reliability
If a classifier skews towards majority class samples, then the average classification cost decreases, but the classification effectiveness for minority class samples deteriorates
Solution Approach 1:
The patent segments the classification problem into multiple independent sub-tasks, each handled by a dedicated sub-classifier. This segmentation ensures that minority class samples in each dimension receive focused attention from their respective sub-classifiers, improving reliability without requiring a single complex multi-class classifier that would be prone to majority class bias.
Solution Approach 2:
The shared feature extraction network provides a universal foundation for all sub-classifiers, enabling them to leverage common facial features while maintaining their specialized focus on specific beauty dimensions. This multi-functional architecture improves minority class classification effectiveness across all dimensions while managing complexity through parameter sharing and feature reuse.
Data Source
AI summary
A multi-dimensional task facial beauty prediction method and system, and a storage medium are disclosed. The method includes the steps of: at a training phase, using first facial images to optimize a shared feature extraction network for extracting shared features and to train a plurality of sub-task networks for performing facial beauty classification tasks; at a testing phase, extracting shared features of second facial images; inputting the shared features to the trained plurality of sub-task networks; and obtaining a first beauty prediction result based on first output results of the plurality of sub-task networks.