修补
人工智能
图像复原
计算机科学
计算机视觉
背景(考古学)
GSM演进的增强数据速率
增采样
块(置换群论)
图像(数学)
失真(音乐)
一致性(知识库)
备份
壁画
相似性(几何)
网络体系结构
保险丝(电气)
迭代重建
图像编辑
边缘增强
图像处理
计算机图形学(图像)
工件(错误)
直方图
深度学习
图像分割
作者
Shengping Xiong,Ying Yu
出处
期刊:
日期:2025-12-06
卷期号:: 1-7
被引量:1
标识
DOI:10.1145/3743093.3770966
摘要
Ancient murals have significant historical, cultural, and artistic value, but have suffered severe damage due to various forms of degradation. Applying digital image inpainting techniques to mural restoration is of great importance. However, traditional and deep learning-based methods still struggle to achieve structural similarity and visual plausibility. To address this, we propose a two-stage restoration framework, MS2G-Net, comprising an Edge and HOG Refinement Network (EHRN) and an Image Refinement Network (IRN). EHRN employs a single-generator dual-discriminator architecture to simultaneously predict Edge and Histograms of Oriented Gradients (HOG) from corrupted inputs. Then, IRN utilizes these predicted structural priors to guide the image restoration process. Specifically, we design an Edge and HOG involved Mask-aware Pixel-shuffle Downsampling (EHMPD) module to hierarchically fuse multi-scale structural features and mitigate information loss during downsampling. To enable both global and local modeling while retaining useful information, we introduce a Spatial-aware Hybrid Attention (SHA) mechanism, integrated with a Gated-Dwconv Feed-forward Network (GDFN) and a Context Broadcasting Feed-forward Network (CBFN) to form a unified "GDFN-SHA-CBFN" structure, termed Spatial-aware Hybrid Attention Block (SHAB). Furthermore, we propose a Structure-Constrained Loss in the HSV color space to enhance color consistency in the restoration results. Extensive qualitative and quantitative experiments demonstrate that our method produces structurally coherent and visually plausible restorations across various mural degradations.
科研通智能强力驱动
Strongly Powered by AbleSci AI