基于CNN-Transformer混合结构的图像去摩尔纹网络  

CNN-Transformer Hybrid Network for Image Demoiréing

作　　者：张祥 吕东辉[1] 冯国瑞[1] Zhang Xiang

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《工业控制计算机》2025年第3期4-6,10,共4页Industrial Control Computer

摘　　要：随着卷积神经网络在计算机视觉领域的发展,现有的去摩尔纹方法对于常规尺寸图像效果良好。然而当图像分辨率变大或者摩尔纹变得复杂时,这些方法表现不佳。为了解决这个问题,提出了一种基于混合CNN-Transformer结构的图像去摩尔纹网络。该网络利用基于移动十字形窗口的Transformer处理高级语义特征,扩大了网络的感受野,从而更好地捕捉远程特征之间的依赖关系;利用纹理感知细化模块处理低级语义特征,更好地重建细节纹理。在公共数据集UHDM、FHDMi和TIP2018上对模型进行评估,证明该模型优于当前主流的方法。With the development of convolutional neural networks in the field of computer vision,existing demoiréing methods perform well on normal size image.However,these methods still face challenges when facing high-resolution images and complex moirépatterns.This paper proposes a novel U-shaped network which employs a hybrid CNN-Transformer structure.The network utilizes Transformer based on shifted cross-shaped window to process high-level features,which expands the receptive field of the network to better capture the dependency between remote features.Texture-aware refinement module is used to process low-level features and reconstruct detailed textures.Experimental results on the public dataset UHDM.FHDMi and TIP2018 indicate that the proposed method significantly outperforms the current mainstream methods.

关 键 词：图像去摩尔纹 混合结构 TRANSFORMER 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]