基于掩码损失的人脸图像生成方法

A Maksed Loss Function for Facial Image Generation

当前主流的基于生成对抗网络(Generative Adversarial Network,GAN)的图像生成方法,在生成真实度较高的人脸图像方面取得了显著进展,但在生成人脸图像的头发、牙齿等细节区域时易出现失真现象。针对存在的问题,提出掩码损失,并将其整合到Style GAN2中。该损失函数通过人脸分割网络生成人脸掩码,基于掩码调整生成图像在细节和非细节区域的贡献程度,以提高细节区域的合成质量。实验结果表明,所提方法显著改善了头发、牙齿等细节区域的合成质量,提高了生成图像的真实度。

Current mainstream image generation methods based on Generative Adversarial Network(GAN)have made significant progress in generating high-fidelity facial images.However,when generating detailed areas such as hair and teeth in facial images,distortion issue arises.To solve this issue,we propose a masked loss and integrate it into StyleGAN2.The masked loss generates facial mask using a facial segmentation network and adjusts the weight of the generated image in the detail and non-detail areas based on this mask,aiming to improve the quality in the detailed areas.Experimental results demonstrate that the proposed method significantly improves the quality of detailed areas such as hair and teeth,thereby enhancing the realism of the generated images.