Facial beauty prediction fusing transfer learning and broad learning system
A new hybrid AI system fuses transfer learning with broad learning to tackle subjective facial analysis.
A research team led by Junying Gan and Xiaoshan Xie has introduced a novel AI architecture designed to tackle the complex challenge of Facial Beauty Prediction (FBP). The core innovation lies in fusing two distinct machine learning paradigms: transfer learning and a Broad Learning System (BLS). The model, named E-BLS, first uses a pre-trained EfficientNet convolutional neural network (CNN) as a feature extractor, leveraging transfer learning to overcome the typical lack of large-scale, labeled facial beauty datasets. These extracted features are then fed into a BLS, a type of flat network known for its rapid training and incremental learning capabilities.
This hybrid approach directly addresses two major FBP hurdles: data scarcity and the complexity of modeling human aesthetic judgment. The team also developed an enhanced version, ER-BLS, which adds a dedicated connection layer between the feature extractor and the BLS for improved information flow. Experimental results demonstrate that both E-BLS and ER-BLS achieve higher accuracy than previous methods that used BLS or CNNs in isolation. The system's effectiveness suggests it could be adapted for other pattern recognition tasks like object detection and image classification, providing a faster, more data-efficient modeling framework.
- Fuses transfer learning (EfficientNet) with a Broad Learning System (BLS) for faster, more accurate model training.
- Achieves higher Facial Beauty Prediction (FBP) accuracy than previous standalone BLS or CNN methods.
- Designed to overcome data scarcity and overfitting in subjective analysis tasks like aesthetic judgment.
Why It Matters
Provides a more efficient AI framework for subjective analysis tasks, with potential applications in personalized aesthetics, marketing, and social robotics.