Enhancing Cross-Age Facial Recognition with T2T-ViT Networks and Multi-Scale Attention Decomposition

Abstract

This paper presents a cross-age facial recognition model that integrates Convolutional Neural Networks (CNN) with Transformers. The model first utilizes a depth-separable T2T-ViT network to extract rich facial features. Subsequently, it employs a multi-scale attention decomposition module to nonlinearly decouple age and identity features. The feature decomposition is jointly constrained by mutual information minimization, cross-entropy, and the Arcface function. The model achieves accuracy rates of 94.97%, 99.51%, and 95.81% on three benchmark datasets: FG-NET, CACD_VS, and CALFW, respectively, matching or surpassing the state-of-the-art (SOTA) performance. These results indicate that the proposed model can extract robust facial information and efficiently decouple features, achieving advanced recognition performance.