摘要
为解决现有算法容易产生图像模糊或纹理失真的问题,提出了一种级联式生成对抗网络图像修复模型.该模型由粗化和优化生成子网络串联而成.在粗化生成网络中设计了一种并行卷积模块,由3层卷积通路和1个深层卷积通路并联组成,当网络层数较深时,可解决梯度消失问题;在深层卷积通路中提出了一种特征提取模块,可利用不同大小的卷积核来获取更加丰富的图像信息.此外,在优化生成网络中提出了一种级联残差模块,通过对4个通道的双层卷积进行交叉级联,可有效增强特征复用;将卷积结果与模块输入特征图的元素对应相加,进行局部残差学习,可提高网络的表达能力;同时采用空洞卷积,可以充分利用上下文信息,保留更多的图像底层细节,实现图像的精细修复.仿真实验结果表明,本文算法修复图像视觉效果好,在3个数据集上峰值信噪比(PSNR)分别为18.4532、18.5496、21.5299;结构相似度(SSIM)为0.8972、0.9683、0.8956,量化结果在对比算法中均为最高,实现复杂结构和纹理信息的自动修复.
To solve the problem of image blur or texture distortion in the existing algorithms,this study proposes a new image inpainting model,called the cascaded generative adversarial networks(C-GAN).The model is cascaded by the coarsening and refinement generation of the sub-networks.In the coarsening generation network,a parallel convolution module is designed to solve the gradient disappearance problem of deep network.It is composed of a three-layer convolution path and a deep one in parallel.In the deep convolution path,a feature extraction module is proposed to achieve a richer image information using convolution kernels of different sizes.Additionally,a cascaded residual module is proposed in the refinement generation network to effectively enhance the feature reuse by crosscascading the double-layer convolution with four channels.Besides,a module input feature map is added to the corresponding elements of the convolution result to improve the expressive ability of the network.Simultaneously,employment of the dilated convolution can fully make use of the context information and retain more rock-bottom image details,which is helpful to achieve a fine restoration.Simulation results demonstrate that the proposed algorithm can achieve better visual effects.For dataset 1,2,and 3,the peak signal-to-noise ratio(PSNR)values are 18.4532,18.5496,and 21.5299 and the structural similarity(SSIM)values are 0.8972,0.9683,and 0.8956 respectively.Highest quantification results are achieved using the comparison algorithm,implying that this algorithm can automatically inpaint some complex structures and texture information.
作者
何凯
刘坤
李宸
马希涛
He Kai;Liu Kun;Li Chen;Ma Xitao(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处
《天津大学学报(自然科学与工程技术版)》
EI
CAS
CSCD
北大核心
2021年第9期917-924,共8页
Journal of Tianjin University:Science and Technology
基金
天津市自然科学基金资助项目(14JCQNJC01500).
关键词
图像修复
生成对抗网络
特征提取模块
残差模块
image inpainting
generative adversarial network
feature extraction module
residual module