BeautyNet: Joint Multiscale CNN and Transfer Learning Method for Unconstrained Facial Beauty Prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Pattern Recognition

Background:

• Facial beauty prediction (FBP) is challenging due to limited discriminative face representations and scarce labeled data.
• Existing methods struggle with unconstrained facial images, necessitating advanced deep learning approaches.
• Fine-grained image classification techniques offer potential for improving feature diversity in FBP.

Purpose of the Study:

• To propose BeautyNet, a novel multiscale deep learning architecture for unconstrained facial beauty prediction.
• To enhance the discriminative power of facial features for more accurate attractiveness assessment.
• To address the challenges of limited training data and computational burden in FBP.

Main Methods:

• A multiscale network architecture is employed to capture diverse facial features at different scales.
• Max-feature-map (MFM) activation function is utilized to reduce computational load and accelerate network convergence.
• Transfer learning strategy is implemented to mitigate overfitting caused by limited labeled facial beauty samples.

Main Results:

• The proposed BeautyNet model demonstrates improved discriminative capabilities for facial features.
• The integration of MFM and transfer learning enhances model performance and efficiency.
• Extensive experiments on the LSFBD dataset show superior performance compared to state-of-the-art methods.

Conclusions:

• BeautyNet effectively addresses key challenges in facial beauty prediction, including data scarcity and feature representation.
• The multiscale architecture combined with MFM and transfer learning offers a robust solution for automated attractiveness assessment.
• The model achieved a classification accuracy of 67.48% on the LSFBD dataset, setting a new benchmark.