联合Swin Transformer和UNet的GAN人脸修复算法
Combined Swin Transformer and UNet for GAN face inpainting
张梦澜1
作者信息
- 1. 太原师范学院计算机科学与技术学院,晋中 030619
- 折叠
摘要
基于GAN的人脸修复技术大都采用CNN进行修复,忽略了人脸修复的全局信息和整体均匀性,从而导致修复结果不理想.基于此问题,提出一种联合Swin Transformer和UNet的GAN人脸修复算法,进行人脸图像修复.该方法整体采用GAN生成器-判别器架构,使用Swin Transformer作为主干网络,用于捕捉图像的全局依赖关系;采用UNet的编码-解码结构,在局部区域进行特征提取和重建.实验结果表明,相较于以往方法,该方法能更好地处理人脸图像修复任务.
Abstract
Most of the GAN-based face restoration techniques use CNN for restoration,ignoring the global information and overall uniformity of face restoration,which leads to unsatisfactory restoration results.Based on this problem,a joint Swin Trans-former and UNet GAN face repair algorithm is proposed for face image repair.The method adopts the GAN generator-discriminator architecture as a whole,using Swin Transformer as the backbone network for capturing the global dependencies of the image,and the encoding-decoding structure of UNet for feature extraction and reconstruction in the local region.The experimental results show that the method can better handle the face image restoration task compared to previous methods.
关键词
生成对抗网络/人脸修复/Swin/Transformer/UNetKey words
generative adversarial networks/face inpainting/Swin Transformer/UNet引用本文复制引用
出版年
2024
NETL
NSTL
万方数据