Multitask Beauty Prediction via Weak Supervision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial beauty prediction technologies relying on deep learning require large training datasets, are prone to overfitting, and incur high costs due to the need for extensive data labeling, especially since real-life tasks often involve multiple interconnected tasks that are difficult to label accurately.
Innovation Solution
A beauty prediction method and device utilizing multitasking and weak supervision, which preprocesses images, allocates them to multiple tasks including a main task and auxiliary tasks, processes features through a feature extraction layer and classification networks (residual and standard neural networks) to reduce dependence on true value labels and enhance generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning networks are used for facial beauty prediction, then prediction accuracy can be improved, but the cost of data labeling increases significantly and the model becomes prone to overfitting
Solution Approach 1:
The patent segments the training data into two distinct parts: clean data (with accurate labels) and noisy data (with unreliable or missing labels). By processing these segments differently through separate processing paths in the neural network, the system can effectively utilize all available data while avoiding the pitfalls of treating all data uniformly, thus reducing labeling costs while maintaining prediction accuracy
Solution Approach 2:
The patent introduces an intermediary mechanism in the form of a multi-task learning framework with shared and task-specific layers. The shared layers learn general features from both clean and noisy data, while task-specific layers refine predictions using clean data. This intermediary structure allows the model to leverage noisy data for general feature learning without allowing it to corrupt the final predictions
2Device complexity
If single-task training is used, then model training is simpler, but the relevance and correlation between multiple real-life tasks are ignored
Solution Approach 1:
The patent implements a multi-task learning architecture where a single neural network model performs multiple related tasks simultaneously. The shared layers of the network learn universal features that are relevant across all tasks, while task-specific branches handle individual task requirements. This allows the model to capture correlations between tasks (such as facial attribute prediction and beauty scoring) while maintaining the ability to optimize for each specific task
Solution Approach 2:
The patent merges multiple task-specific processing branches into a unified multi-task learning framework. By combining the loss functions of different tasks and sharing computational resources across tasks, the system achieves better generalization and utilizes the interrelationships between tasks, improving overall performance while keeping the training process manageable through unified optimization
Data Source
AI summary
A beauty prediction method and device based on multitasking and weak supervision, and a storage medium are disclosed. The method includes the steps of pre-processing inputted facial images; allocating the pre-processed images to multiple tasks; extracting shared image features; and obtaining a plurality of classification results via a plurality of classification networks each including a residual network, a standard neural network and a classifier.