Comprehensive Style Transfer for Facial Images Using Enhanced Feature Attribution in Generative Adversarial Nets

초록

Image-to-image translation is a fundamental task in computer vision that transforms images between domains while preserving essential content. Although adaptive instance normalization (AdaIN) is widely used for style transfer, its reliance on simple statistical measures (mean and variance) may limit its ability to capture complex style characteristics. We propose a novel framework that enhances style transfer by combining AdaIN with Gram matrices, leveraging the complementary strengths of both approaches. Our method introduces two key innovations for enhanced feature attribution: 1) dual Gram matrix-based loss functions (G1 and G2), which operate at different stages of the generation process to capture richer style information by establishing deeper correlations between feature maps, and 2) a balanced training objective that integrates perceptual loss with cycle-consistency loss to maintain content fidelity during style transfer. This comprehensive feature attribution mechanism enables our model to decompose and reassign stylistic elements across domains more precisely. Through ablation studies, we demonstrate that each component of our framework contributes to performance improvements, with the complete model achieving the best results on both the CelebA-HQ and FFHQ datasets. Our comprehensive evaluation, using distribution similarity metrics, classification-based assessments, and visual comparisons, demonstrates that our approach effectively captures and transfers complex style characteristics while preserving content integrity, outperforming state-of-the-art models. Specifically, our model achieves superior Fr & eacute;chet Inception Distance (FID) scores (19.88 vs. 24.22) and recognition accuracy (0.966 vs. 0.941) compared to StarGAN v2, confirming the performance gains introduced by our enhanced feature attribution strategy.

키워드