基于上下文语义递阶推理的图像盲修复方法研究

Research on Image Blind Inpainting Algorithm Based on Contextual Semantic Recursive Reasoning

现有的基于深度学习的图像修复方法通常需提供成对的受损图像和掩码图像作为输入,而现实中破损图像对应的掩码通常难以获得,为此,提出一种基于上下文语义递阶推理的图像盲修复网络。该网络由局部填补模块和细节优化模块组成,前者根据图像局部上下文语义自动估计图像中的污损区域,并完成初步的像素填充;而后者结合局部区域的多尺度特征,修复前一阶段输出图像中语义不和谐的区域,最终生成清晰自然的图像。结果表明:该方法性能优于同类算法,且能生成全局语义一致的清晰图像。

Existing image inpainting methods based on deep learning mostly require pair inputs of damaged image and binary mask.However,mask image of damaged image is not available in practice.In this work,a blind inpainting method based on contextual semantic recursive reasoning is proposed.The whole framework comprises of two modules named as local padding network(LPN)and detail refining network(DRN).Under the guidance of local contextual semantic,dirtied regions in image is estimated automatically and coarsely filled by LPN.Then image patches of inconsistent semantic in coarse image resulted from LPN is fixed by DRN,which uses multi-scale feature of local context,and a clear and natural image is obtained consequently.Extensive experiments show that the presented algorithm performs surpass the counterparts and can produces clear images of higher global semantic consistency.