损失自适应的高感知质量生成对抗超分辨率网络  

Loss adaptive GAN for great perceptual image super-resolution

作　　者：林旭锋 吴丽君[1] 陈志聪[1] 林培杰[1] 程树英[1] LIN Xufeng;WU Lijun;CHEN Zhicong;LIN Peijie;CHENG Shuying(College of Physics and Information Engineering,Fuzhou University,Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350108

出　　处：《福州大学学报(自然科学版)》2025年第1期26-34,共9页Journal of Fuzhou University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62271151);福建省自然科学基金资助项目(2021J01580)。

摘　　要：为解决生成对抗网络训练过程中因损失简单加权导致的图像感知质量下降问题,提出损失自适应调整的生成对抗超分辨率网络(LA-GAN).首先,该方法设计通过计算角点分布的相关强度大小,区分规则纹理区域与不规则纹理区域.其次,基于不同区域,设计了区域自适应生成对抗学习框架.在该框架中,网络只在不规则纹理区域中进行对抗学习,提高感知质量.此外,基于下采样图像和图像块相似性的重组图像取代训练集中的高分辨率图像,实现平均绝对损失在不规则纹理区域弱约束网络,在规则纹理区域强约束网络,保证图像信号保真度.最后,通过实验证明经过优化的网络在信号保真度和感知质量方面皆有提升.In order to solve the degradation of image perception quality due to the simple weighting of losses during the training process of generative adversarial networks,this paper proposes the loss adaptive generative adversarial super-resolution network(LA-GAN).Firstly,the method is designed to distinguish regular texture regions from irregular texture regions by calculating the correlation intensity magnitude of the corner point distribution.Secondly,a region-adaptive generative adversarial learning framework is designed based on different regions.In this framework,the network performs adversarial learning only in the irregular texture region to improve the perception quality.Moreover,instead of the high-resolution ground truth image,a patch-wise recombined images based on down sampling and image block similarity to achieve mean absolute error that weakly constraints in irregular texture regions and strongly constraints in regular texture regions to ensure the signal fidelity.Finally,experimental results demonstrate that the proposed LA-GAN can effectively improve both signal fidelity and perceptual quality.

关 键 词：超分辨率 生成对抗网络 损失函数 区域自适应 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]