Face Attribute Transfer Fusing Feature Enhancement and Structural Diversity Loss Function

Abstract

In the process of face attribute transfer, non-frontal and occluded face images often suffer from low generation quality, missing facial edges, and a lack of diversity. To address these challenges, we present the FES-StarGANv2, an unsupervised multi-domain face attribute transfer network. In the feature extraction phase, we incorporate an attention-guided feature fusion module aimed at enhancing image details while preserving the overall integrity of the transferred images. Moreover, a style code extraction module is presented, refining the style code of the target domain, enhancing the learning capabilities of the generator. To further augment image diversity and authenticity, a face image optimization module and a structural diversity loss function are integrated. Experimental results reveal that, in comparison with the baseline StarGANv2, our approach achieves substantial improvements of 23% and 3.9% in FID and LPIPS metrics, respectively, attaining optimal 13 and 0.453. Notably, in terms of visual quality, significant enhancements were observed, particularly in addressing issues of low image quality and edge deficiencies. The FES-StarGANv2 approach effectively addresses the challenges associated with non-frontal and occluded facial images.