江西科学2025，Vol.43Issue(4)：643-650,8.DOI:10.13990/j.issn1001-3679.2025.04.010
基于GAM注意力机制和区域归一化的三阶段网络图像修复
Three-Stage Network for Image Inpainting Based on GAM Attention Mechanism and Region Normalization
摘要
Abstract
To address the issues of blurriness and artifacts in existing image inpainting meth-ods,this paper proposes a three-stage generative network model for image inpainting that incorporates a Global Attention Mechanism(GAM)and Region Normalization(RN).The GAM module is inserted before the decoder to model the correlation between contextual information and missing regions,thereby capturing a global receptive field.A learnable re-gion normalization module(RN_L)is embedded within the decoder to extract effective fea-tures from the incomplete image.Additionally,the activation function of the residual block is changed from ReLU to SELU,which effectively reduces information retention.Experi-ments conducted on the Places2 and Paris Street View datasets demonstrate that both quanti-tative and qualitative analysis of the proposed model are superior to those of existing meth-ods.The repaired images exhibit structures and texture details more similar to the real ima-ges,achieving improved visual performance and significantly mitigating blurring and artifact problems in image inpainting.关键词
图像修复/注意力机制/三阶段生成网络Key words
image inpainting/attention mechanism/three-stage generation network分类
信息技术与安全科学引用本文复制引用
吉子嘉,章银娥..基于GAM注意力机制和区域归一化的三阶段网络图像修复[J].江西科学,2025,43(4):643-650,8.基金项目
国家自然科学基金项目(62041210) （62041210）
江西省人文社会科学规划项目(JC22118) （JC22118）
江西省教育科学"十四五"规划项目(23YB172) （23YB172）
江西省教育厅科技项目(GJJ2201206). （GJJ2201206）
