改进型轻量级GAN的红外图像超分辨率方法  

Infrared Image Super-resolution via Improved Lightweight GAN

作　　者：胡德敏[1] 闵天悦 HU De-min;MIN Tian-yue(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2022年第8期1711-1717,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61170277,61472256)资助;上海市教委科研创新重点项目(12zz137)资助;上海市一流学科建设项目(S1201YLXK)资助.

摘　　要：红外图像超分辨率是图像超分辨率重建的子领域,基于深度学习的方法侧重于研究色彩和纹理丰富的RGB图像重建,对于像素分布均匀、对比度低、高频细节特征丢失的红外图像提取特征效率低.本文采用生成对抗网络(GAN)针对红外图像提出了一种能重建细节纹理超分辨率方法,用轻量级注意力残差块(Lightweight attention residual block,LARB)构建生成器网络,以低成本、高效率提取到红外图像的像素特征信息;结合特征激活前的感知损失、Huber损失和Wasserstein距离使模型稳定收敛,减少图像重建后伪影的产生;引入近红外图像数据集与红外特征图线性灰度变换使模型学习更多纹理特征以修复高频细节.实验结果显示,在PSNR的比较中,本文的模型在生成器参数(Params)仅有542K情况下大幅领先于参数为1518K的SRGAN;在部分测试数据集中SSIM高于参数为16697K的ESRGAN,表明了方法的有效性.Infrared image super-resolution is a field of image super-resolution reconstruction,and methods via deep learning tend to focus on RGB images with rich colors and textures,which are inefficient for extracting features from infrared images with uniform pixel distribution,low contrast,and loss of high-frequency details.This paper proposes a super-resolution method based on generative adversarial networks(GAN)that can reconstruct detailed textures for infrared images.Its generator is constructed by a series of Lightweight attention residual blocks(LARB),which can extract pixel feature information at low cost and high efficiency from infrared images.It combines with perception loss before feature activation,Huber loss and Wasserstein distance to converge stably and reduce artifacts after image reconstruction.We use near infrared image datasets and linear grayscale transform of infrared feature maps to help our model learn more texture features.The experimental results show that in the PSNR comparison,our model generator with a parameter of 542K is significantly ahead of SRGAN with a parameter of 1518K,and is close to ESRGAN with a parameter of 16697K.In the SSIM comparison,some test results are higher than ESRGAN,which indicating the effectiveness of our method.

关 键 词：红外图像 轻量 生成对抗网络 超分辨率 注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]