现代电子技术2025，Vol.48Issue(13)：43-49,7.DOI:10.16652/j.issn.1004-373x.2025.13.006
基于注意力机制和ACT网络的人脸图像修复
Face image inpainting based on attention mechanism and ACT network
摘要
Abstract
A face image inpainting method based on convolutional block attention module(CBAM)and aggregated contextual transformation(ACT)network is proposed to make the completion of missing semantic features in facial images more realistic and to enhance the recovery of detailed information.In this method,the two branches of the baseline model are retained.In the semantic and image filtering branches,the CBAM layers are added to capture the critical detail information for filling the missing areas in the images.The baseline residual blocks are replaced with ACT residual blocks,which can preserve the rich details outside the missing areas and capture abundant contextual information.This made the semantic information filling in this branch more accurate,effectively removed artifacts and enriched image details.In the kernel prediction branch,the two modules are added to enhance the receptive field and contextual reasoning perception when extracting image features,making the dynamic prediction of filtering kernels more precise.This method was validated on the CelebA-HQ dataset,showing improvements in quantitative metrics such as PSNR(peak signal-to-noise ratio),SSIM(structural similarity index measure)and L1.The qualitative repair results are also clearer and more natural.The study confirms that the proposed method have good effectiveness for facial image inpainting.关键词
图像修复/CBAM注意力机制/ACT网络/编码器-解码器/人脸图像修复/图像滤波Key words
image inpainting/CBAM attention mechanism/ACT network/encoder-decoder/facial image inpainting/image filtering分类
电子信息工程引用本文复制引用
滕林,张乾,徐开丽..基于注意力机制和ACT网络的人脸图像修复[J].现代电子技术,2025,48(13):43-49,7.基金项目
贵州省高等学校智能算法与智能软件协同创新团队(黔教技[2023]061号) （黔教技[2023]061号）
贵州省高等学校大数据分析与智能计算重点实验室(黔教技[2023]012号) （黔教技[2023]012号）
贵州民族大学校级科研项目(GZMUZK[2023]QN10) （GZMUZK[2023]QN10）
贵州民族大学校级科研项目(GZMUZK[2021]YB23) （GZMUZK[2021]YB23）
