融合高效注意力的生成对抗网络图像修复算法  

Generative Adversarial Network Image Restoration Algorithm with Efficient Attention

作　　者：袁雪丰 杜洪波[1] 朱立军[2] 刘雪莉 YUAN Xuefeng;DU Hongbo;ZHU Lijun;LIU Xueli(School of Science,Shenyang University of Technology,Shenyang 110870,Liaoning Province;School of Computer Science and Engineering,North Minzu University,Yinchuan 750021,Ningxia)

机构地区：[1]沈阳工业大学理学院,辽宁沈阳110870 [2]北方民族大学信息与计算科学学院,宁夏银川750021

出　　处：《沈阳工程学院学报(自然科学版)》2025年第1期63-70,共8页Journal of Shenyang Institute of Engineering:Natural Science

基　　金：国家自然科学基金(11861003);辽宁省教育厅高等学校基本科研项目(LJKZ0157)。

摘　　要：针对现有图像修复算法存在细节纹理结构还原效果不佳及修复区域与图像未缺损区域的视觉连通性较差的问题,提出了一种以双判别生成对抗网络为框架,融合高效通道注意力(ECA)和感知损失的图像修复算法(EPGAN)。ECA模块用于调节不同通道的权重,提高特征利用率来获取待修复区域与未缺损区域关联性更高的像素信息,利用VGG16模型提取特征得到感知损失用来学习语义特征差异,消除动态模糊,使修复结果保留更多细节和边缘信息。EPGAN算法在数据集CelebA和Places2上分别做了定性、定量分析及消融实验,根据峰值信噪比和结构相似性的评价结果及修复效果显示:EPGAN算法优于GLCIC、GC和MFCL图像修复算法,且通道注意力和感知损失有效地改善了模型的修复效果。A novel image restoration algorithm(EPGAN)is proposed in this study to address the issues of poor restoration of fine-texture structures and weak visual connectivity between the repaired and undamaged regions,which are commonly observed in existing image restoration algorithms.The proposed algorithm is based on a dual-discriminator generative adversarial network framework,integrating the Efficient Channel Attention(ECA)mechanism and perceptual loss.The ECA module is employed to adjust the weights of different channels,improving the utilization of features to obtain pixel information with higher correlation between the targeted restoration region and the undamaged region.The VGG16 model is utilized to extract features for capturing semantic feature differences and eliminating dynamic blurring,thereby preserving more details and edge information in the restoration results.Qualitative and quantitative analyses,as well as ablation experiments,are conducted on the CelebA and Places2 datasets to evaluate the performance of the EPGAN algorithm.Evaluation results based on peak signal-to-noise ratio and structural similarity demonstrate the superiority of the EPGAN algorithm over CE,GLCIC and MFCL image restoration algorithms.Furthermore,the integration of channel attention and perceptual loss significantly improves the restoration performance of the model.

关 键 词：图像修复 生成对抗网络 通道注意力 感知损失 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]