动态注意力语义特征增强的人脸风格迁移方法研究

Research on Facial Style Transfer Method for Dynamic Attention Semantic Feature Enhancement

摘要: 为了解决基于StyleGAN2网络的人脸风格迁移算法存在的风格迁移效果不明显、不准确，局部纹理不合理等问题。本文提出动态注意力语义特征增强的人脸风格迁移方法。本文提出归一化傅里叶卷积将人脸特征转化到频域中进行处理，使网络更好地捕捉图像纹理、细节和结构等频率信息，提出了一个新的动态注意力结构控制块使生成器更高效地学习风格图像的结构特征并较好保留原始人脸特征；此外，还引入色彩一致性损失函数，该损失函数可以有效保持生成图像和风格图像的颜色相似性。实验结果表明，本文提出的网络在定性和定量实验研究中与最先进的方法相比都取得了最好得分。本文方法在保留内容图像语义信息和细节特征的同时有效学习了风格图像的特征，更好地实现了人脸风格迁移准确性和艺术性。

Abstract: To address the issues of unclear and inaccurate style transfer effects, as well as unreasonable local textures in the StyleGAN2-based face style transfer algorithm, this paper introduces a dynamic attention semantic feature enhancement approach. The normalization Fourier convolution is proposed to process facial features in the frequency domain, allowing the network to more effectively capture image texture, detail, and structural frequencies. A new dynamic attention control block is introduced to enable the generator to more efficiently learn the structural features of style images and better preserve the original facial features. Additionally, a color consistency loss function is introduced to ensure the color similarity between the generated and style images. Experimental results demonstrate that our network achieves the highest scores in both qualitative and quantitative evaluations compared to the leading methods. This approach effectively learns the features of style images while preserving the semantic information and detailed features of content images, enhancing the accuracy and artistry of face style transfer.