融合残差密集与生成对抗网络的红外巡检图像超分辨率重建  

Super-Resolution Reconstruction of Infrared Inspection Images by Integrating Residual Dense and Generative Adversarial Networks

作　　者：刘志坚[1] 陶韵旭 刘航 罗灵琳[1] 张德春 何蔚 LIU Zhijian;TAO Yunxu;LIU Hang;LUO Lingin;ZHANG Dechun;HE Wei(Faculty of Electric Power Engineering,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学电力工程学院,云南昆明650500

出　　处：《昆明理工大学学报（自然科学版）》2023年第5期120-129,共10页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：云南省教育厅科学研究基金项目(2022J0052);云南省基础研究计划青年项目(202201AU070086);昆明理工大学自然科学研究基金资助项目(KKZ3202004042)。

摘　　要：为了提升电力设备红外巡检图像质量,最大程度还原图像内容的有效性和准确性,提出了一种融合残差密集与生成对抗网络的红外图像超分辨率重建方法.将残差密集网络引入到WGAN(Wasserstein Generative Adversarial Networks)网络生成器,使其在训练过程中形成连续记忆机制,提高网络对图像特征的融合能力;进一步,使用谱归一化方法优化WGAN的对抗器网络参数,提升对抗训练的稳定性和效率;构造由对抗、像素、感知和纹理损失构成的综合损失函数,完成生成图像高频细节信息的重建.超分辨率实验结果表明,重建后的电力设备红外巡检超分辨率图像在峰值信噪比上提升至32.048 2 dB,在结构相似性上提升至0.921 4,且视觉效果良好,验证了所提方法能够有效提升图像质量并具备较好的工程应用价值.In order to improve the quality of infrared inspection images for power equipment and maximize the effectiveness and accuracy of image content restoration,a method for super-resolution reconstruction of infrared images is proposed by integrating residual dense and generative adversarial networks.The residual dense network is introduced into the generator of the Wasserstein Generative Adversarial Networks(WGAN) to form a continuous memory mechanism during the training process,enhancing the network's ability to integrate image features.Furthermore,a spectral normalization method is used to optimize the parameters of the adversarial network in WGAN,improving the stability and efficiency of adversarial training.A comprehensive loss function consisting of adversarial,pixel,perceptual,and texture losses is constructed to reconstruct the high-frequency details of the generated images.The experimental results of super-resolution demonstrate that the reconstructed infrared inspection images of power equipment achieve a peak signal-to-noise ratio of 32.048 2 dB and a structural similarity index of 0.921 4,showing good visual quality.This verifies that the proposed method effectively improves image quality and possesses significant engineering application value.

关 键 词：红外巡检图像 超分辨率重建 残差密集网络 谱归一化 多损失融合技术 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TM50[自动化与计算机技术—计算机科学与技术]