基于密集连接注意力块的双生成器图像修复算法  

Dual Generator Image Inpainting Algorithm Based on Densely Connected Attention Block

作　　者：胡海燕 李硕[2] 刘斌[2] HU Haiyan;LI Shuo;LIU Bin(Yulin Electric Power Supply Company,State Grid Shaanxi Electric Power Supply Company,Yulin 719000,China;School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi’an 710021,China)

机构地区：[1]国网陕西省电力有限公司榆林供电公司,陕西榆林719000 [2]陕西科技大学,电子信息与人工智能学院,陕西西安710021

出　　处：《微型电脑应用》2024年第2期1-5,共5页Microcomputer Applications

基　　金：国家自然科学基金项目(61871260)。

摘　　要：针对图像修复痕迹明显、模型训练不稳定等问题,设计一种结合密集连接注意力块的图像修复算法。在生成器中引入精修复和粗修复二阶段修复网络,并在精修复网络中使用4个通道注意力块设计的密集连接注意力块;同时,增设VGG16特征提取模型,引入WGAN-GP作为判别器损失函数,以多损失融合的方式提高图像的修复效果。在CelebA数据集上验证模型的修复效果,该算法在主客观指标上均优于DCGAN、CE和DD这3种主流算法。An image inpainting algorithm is designed by combining densely connected attention blocks for the problems of obvious image inpainting marks and unstable model training.A two-stage inpainting network with fine inpainting and coarse inpainting is introduced into the generator,and a densely connected attention block with four channel attention blocks is used in the fine inpainting network.At the same time,the VGG16 feature extraction model is added,and WGAN-GP is introduced as the discriminator loss function to improve the image inpainting effect by multi-loss fusion.To verify the inpainting effect of the model on CelebA dataset,the algorithm outperformed three mainstream algorithms,DCGAN,CE,and DD,in both subjective and objective indicators.

关 键 词：图像修复 生成对抗网络 通道注意力块 密集连接网络 VGG16 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]