Identity-Preserving Face Aging With Multi-Attribution Fusion and Multi-Scale Attention

Despite the recent advances of face age transformation in the age accuracy of face synthesis or the integrity of identity information, only age used as the condition of aging is insufficient to depict the real aging process. In this work, we propose a novel GAN-based identity-preserving aging method with Multi-scale attention and Multi-attribution fusion (named M2aging) to address this problem. In particular, we design a High-fidelity Multi-scale Attention(HiMA) module to capture minute and local features at different scales, improving the model's identity-preserving ability. In addition, a Multi-attribution Fusion Modulator (MFM) is proposed by fusing age, gender, and race attributions to control the aging process, which enhances attribution preservation without introducing any additional classifiers and training losses. Experimental results show the proposed M2aging outperforms other SOTA techniques in terms of identity preservation and accurate age conversion. The ablation experiments manifest that HiMA can effectively heighten the fidelity of the transformed images, and MFM can effectively control the aging process according to age, gender, and race.