摘要
针对受损区域修复存在语义不连贯,纹理不清晰的问题,本文提出了一种基于改进注意力机制的生成对抗网络图像修复方法。用U-Net作为生成器主干,为充分学习图像特征,提高编码器中特征的利用率,在编码阶段引入改进的通道注意力模块,同时为了克服长距离对于信息的依赖,在跳跃连接层中添加转移连接层,保持图像信息的连贯性。此外在原本的重构损失函数和对抗损失函数中,添加指导损失函数与风格损失函数,增加了整个网络的稳定性。实验结果表明在Celeb A和Places2数据集上,本文方法取得了较好的修复效果。
Aiming at the problems of semantic incoherence and unclear texture in damaged area repair,a generative adversarial network images restoration method based on improved attention mechanism is proposed in this paper.U-net is used as the backbone of the generator.In order to fully learn images features and improve the utilization of features in the encoder,an improved channel attention module is introduced in the coding stage.At the same time,for the purpose of overcoming the dependence on information over a long distance,a transfer connection layer is added to the jump connection layer to maintain the coherence of images information.In addition,in the original reconstruction loss function and countermeasure loss function,guidance loss function and style loss function are added to increase the stability of the whole network.The experimental results show that the proposed method has achieved good restoration results on Celeb A and Places2 data sets.
作者
张剑飞
张洒
夏万贵
ZHANG Jianfei;ZHANG Sa;XIA Wangui(School of Computer&Information Engineering,Heilongjiang University of Science&Technology,Harbin 150022,China)
出处
《智能计算机与应用》
2022年第6期141-145,共5页
Intelligent Computer and Applications
基金
国家自然科学基金(61803148)
关键词
图像修复
通道注意力
转移连接层
生成对抗网络
images restoration
channel attention
transfer connection layer
generative adversarial network
作者简介
张剑飞(1978-),女,博士,副教授,主要研究方向:人工智能、计算机视觉;张洒(1993-),女,硕士研究生,主要研究方向:计算机视觉;夏万贵(1993-),男,硕士研究生,主要研究方向:人工智能。
