摘要
针对当下人脸去手势遮挡任务中常出现的结构缺失和纹理模糊等问题,文章提出一种基于边缘条件和注意力机制的两阶段修复网络——EmmNet。第一阶段网络为第二阶段细节修复提供边缘指导信息,以避免出现过度平滑等问题。第二阶段网络中的并行多扩张卷积模块可在有效扩大网络感受野的同时提高对有效像素的利用率。此外,注意力模块可促使网络生成具有全局一致性,使研究者获得符合原图特征的修复图像。实验结果表明,EmmNet在去手势遮挡任务中可以生成轮廓结构更加完整流畅,细节纹理更加清晰自然的人脸图像。
Aiming at the problems of lack of structure and blurred texture that often appear in the current face removing gesture occlusion task,this paper proposes a two-stage inpainting network based on edge conditions and attention mechanism—EmmNet.The onestage network provides edge guidance information for the second-stage detail restoration to avoid problems such as over-smoothing.The parallel multi-expansion convolution module in the second-stage network can effectively expand the network receptive field and improve the utilization of effective pixels.In addition,the attention module can promote global consistency in network generation,enabling researchers to obtain repaired images that match the original image features.The experimental results show that EmmNet can generate a face image with a more complete and smooth outline structure and clear and natural detail texture in the removing gesture occlusion task.
作者
欧静
文志诚
OU Jing;WEN Zhicheng(Hunan University of Technology,Zhuzhou 412007,China)
出处
《现代信息科技》
2023年第11期97-100,104,共5页
Modern Information Technology
关键词
卷积神经网络
生成对抗网络
人脸修复
注意力机制
Convolutional Neural Network
Generative Adversarial Network
face restoration
attention mechanism