华中科技大学学报（自然科学版）2025，Vol.53Issue(5)：178-184,7.DOI:10.13245/j.hust.250374
由粗到细的内容一致性图像修复方法
Coarse to fine approach to content-consistent image inpainting
摘要
Abstract
To solve the problem that it was unable to take into account the global structure and texture details when repairing large areas of missing images,a coarse-to-fine content-consistent image inpainting network was proposed.First,the broken image was fed into the coarse network,and the codec structure used gated convolution and deconvolution for image extraction and recovery of features from damaged images.Then,the damaged content was reconstructed through the decoder,and the coarse repair result was output in combination with the original image.The second stage of the local refinement network enhanced the feature extraction and spatial information processing in the detail part through the combination of convolutional layers,normalisation layers,ReAt structure,and deconvolutional layers.The global refinement network of the third stage of the attention mechanism was based on the U-Net architecture,and the coherent semantic attention(CSA)module was integrated with the Condensed Attention module.Feature extraction and spatial information processing were optimized through symmetry,hopping connectivity and multilevel attention mechanisms,which could significantly improve the accuracy,coherence and quality of image inpainting.Through evaluations on three publicly available datasets,results show that superior performance in large-mask restoration and content consistency is demonstrated by the proposed algorithm,and both subjective and objective assessments confirm its outperformance over the comparative methods.关键词
图像修复/U-Net架构/注意力机制/生成对抗网络/多阶段修复Key words
image inpainting/U-Net architecture/attentional mechanism/generative adversarial networks/multi-stage inpainting分类
计算机与自动化引用本文复制引用
魏赟,王璐璐,辛子昊,邬开俊..由粗到细的内容一致性图像修复方法[J].华中科技大学学报（自然科学版）,2025,53(5):178-184,7.基金项目
甘肃省自然科学基金资助项目(23JRRA913) （23JRRA913）
内蒙古自治区重点研发与成果转化计划资助项目(2023YFSH0043) （2023YFSH0043）
甘肃省教育厅高校教师创新基金资助项目(2025A-054). （2025A-054）
