CILF-CIAE: CLIP-driven image-language fusion for correcting inverse age estimation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial age estimation is crucial for applications like age verification and access control.
• Existing Transformer models for age estimation face high memory demands due to quadratic complexity.
• Contrastive Language-Image Pre-training (CLIP) shows promise but its integration with error feedback for age estimation is unexplored.

Purpose of the Study:

• To propose a novel CLIP-driven Image-Language Fusion for Correcting Inverse Age Estimation (CILF-CIAE).
• To address memory inefficiency in Transformer-based age estimation.
• To enhance the accuracy and robustness of age estimation systems.

Main Methods:

• Utilized CLIP for extracting aligned image and text features.
• Developed FourierFormer, a linear log complexity Transformer architecture for efficient image-text fusion.
• Implemented contrastive multimodal learning to bridge the semantic gap between modalities.
• Introduced reversible age estimation with end-to-end error feedback.

Main Results:

• CILF-CIAE demonstrated superior performance across six benchmark datasets.
• Achieved a Mean Absolute Error (MAE) of 1.68 on MORPH-S2, outperforming LRA-GNN (2.21) and MCGRL (1.77).
• The proposed FourierFormer offers significant memory savings compared to traditional attention mechanisms.

Conclusions:

• CILF-CIAE effectively integrates CLIP and an efficient Transformer for robust age estimation.
• The method significantly reduces age prediction errors through advanced fusion and feedback techniques.
• The findings highlight the potential of CILF-CIAE for real-world age estimation applications.