Multi-Task Facial Beauty Prediction via Graph-Based Task Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial beauty prediction technologies using deep learning suffer from increased redundancy and training burden due to single-task learning and unnecessary combinations in multi-task learning, affecting efficiency and precision of classification and recognition.
Innovation Solution
A facial beauty prediction method and device based on multi-task migration, which performs similarity measurement using a graph structure to find an optimal combination of tasks, constructs a facial beauty prediction model with a feature sharing layer, migrates feature parameters from an existing large-scale facial image network, and includes pre-processing, independent feature extraction, and classification layers to reduce redundancy and improve training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multi-task learning is applied to facial beauty prediction, then the comprehensiveness of facial beauty factor recognition is improved, but the redundancy of deep learning tasks and training burden increase
Solution Approach 1:
The patent segments the multi-task learning process into distinct components: a feature extraction network that processes images once, and multiple task-specific prediction networks (beauty score prediction, facial feature point detection, expression recognition, gender recognition, age recognition) that operate independently. This segmentation eliminates redundant feature extraction for each task while maintaining comprehensive recognition capabilities.
Solution Approach 2:
The patent implements a universal feature extraction network that serves all five prediction tasks simultaneously. The shared network extracts common facial features once, and these features are then utilized by all task-specific prediction networks, reducing overall computational redundancy while maintaining comprehensive task coverage.
2Productivity
If single-task learning is used for facial beauty prediction, then the training burden is reduced, but the correlation between tasks is ignored
Solution Approach 1:
The patent merges multiple single-task networks into a unified multi-task framework where the feature extraction network is shared across all tasks. This combination preserves the training efficiency of single-task learning by avoiding redundant feature extraction, while simultaneously capturing task correlations through the shared representations in the common network.
Data Source
AI summary
Disclosed are a facial beauty prediction method and device based on multi-task migration. The method includes: performing similarity measurement based on a graph structure on a plurality of tasks to obtain an optimal combination of the plurality of tasks; constructing a facial beauty prediction model including a feature sharing layer based on the optimal combination; migrating feature parameters of an existing large-scale facial image network to the feature sharing layer of the facial beauty prediction model; inputting facial images for training to pre-train the facial beauty prediction model; and inputting a facial image to be tested to the trained facial beauty prediction model to obtain facial recognition results.