一种基于自注意力机制的人脸图像补全算法  

A FACE IMAGE INPAINTING ALGORITHM BASED ON SELF ATTENTION MECHANISM

作　　者：杨博文 何衡湘[1] 邓洪峰[1] Yang Bowen;He Hengxiang;Deng Hongfeng(Southwest Institute of Technical Physics,Chengdu 610000,Sichuan,China)

机构地区：[1]西南技术物理研究所,四川成都610000

出　　处：《计算机应用与软件》2024年第8期266-270,318,共6页Computer Applications and Software

摘　　要：针对目前深度学习的方法在大面积信息缺失的人脸图像进行补全应用中,补全结果出现纹理细节模糊、结构变形扭曲等问题,提出一种基于自注意力机制的图像补全算法。该算法将待补全的图像输入基于跳跃连接的粗生成网络,得到初步修复;将初步结果输入自注意力感知分支和混合空洞卷积分支共同编码,再通过解码得到生成结果;由双判别器完成判别优化工作。通过人脸图像CelebA-HQ数据集进行实验与测试,所提方法的补全结果在客观和主观评价方面,优于deepfill和PLC两种算法。In the application of current deep learning methods in large area information missing face image inpainting,the inpainting results show issues such as blurred texture details,structural deformation,and distortion.Aimed at these problems,an image inpainting algorithm based on self-attention mechanism is proposed.The image to be completed was input into the rough generation network based on skip-connection to get the preliminary repair.The initial results were input into the self-attention sensing branch and the hybrid hole convolution branch to encode together,and the generated results were obtained by decoding.The dual discriminant was used to optimize the discriminant.Through the experiments and tests on face image CelebA-HQ dataset,the results show that the proposed method has better inpainting effect than the deep fill and PLC algorithms in objective and subjective evaluation.

关 键 词：图像补全 生成对抗网络 跳跃连接 自注意力机制 混合空洞卷积 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]