基于级联残差对抗生成网络的超分辨重建  

Super resolution based on cascading residual generative adversarial network

作　　者：祁成晓 刘芳[2] 孙策 曲振方 朱福珍[1] QI Chengxiao;LIU Fang;SUN Ce;QU Zhenfang;ZHU Fuzhen(College of Electrical Engineering,Heilongjiang University,Harbin 150080,China;School of Information Engineering,East University of Heilongjiang,Harbin 150086,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080 [2]黑龙江东方学院信息工程学院,哈尔滨150086

出　　处：《黑龙江大学自然科学学报》2022年第3期365-371,共7页Journal of Natural Science of Heilongjiang University

基　　金：国家自然科学基金资助项目(61601174);黑龙江省博士后科研启动金项目(LBH-Q17150);黑龙江省普通高等学校电子工程重点实验室(黑龙江大学)开放课题资助及省高校科技创新团队资助项目(2012TD007);黑龙江省省属高等学校基本科研业务费基础研究项目(KJCXZD201703);黑龙江省自然科学基金资助项目(F2018026)。

摘　　要：图像超分辨率技术在遥感领域的应用越来越多,卷积神经网络(Convolutional neural network, CNN)在图像超分辨率任务方面取得了比传统方法更显著的改进。为了解决超分辨重建算法重建图像细节不清晰的问题,对基于对抗生成网络的超分辨重建算法进行了研究,以恢复接近人眼感知质量的超分辨重建图像。对抗生成网络分为生成器子模块和判别器子模块,生成器模块提取的低频图像特征在进入残差块前进行分组卷积,达到了降低参数量的目的。每3个残差块构成一个级联块,级联块通过级联的方式聚合不同级联块之间图像的特征,以实现信息流传递到更加深层的网络中,经过上采样完成生成器中的图像重建过程。判别器网络用于判别是否为生成器重建图像和真实的高分辨率图像,在真实高分辨率图像的判别过程中提高模型能力。同时,研究了加入感知损失对超分辨重建图像的影响,使重建图片纹理细节更加丰富。实验结果表明,重建图像的峰值信噪比较原来算法提高了0.48 dB,结构相似度提高了0.023,该模型在主观视觉评价和客观量化评估上有所提升。Remote sensing image super-resolution techniques are increasingly used in the field of remote sensing.The Convolutional neural network(CNN)has achieved significant improvements over traditional methods for image super-resolution tasks.In order to solve the problem of unclear details in the reconstructed images of current super-resolution reconstruction algorithms,super-resolution reconstruction algorithms based on adversarial generative networks are investigated as a means to recover super-resolution reconstructed images that are closer to the perceptual quality of the human eye.The adversarial generative network is divided into a generator sub-module and a discriminator sub-module.The low-frequency image features extracted by the generator module are grouped and convolved before entering the residual blocks,so as to achieve the purpose of reducing the number of parameters.Every three residual blocks form a cascading block,and the cascading blocks aggregate the features of the images between the different cascading blocks in a cascading manner,so that the information flow can be passed to the deeper network,and then up-sampled to complete the image reconstruction process in the generator.The discriminator network is used to discriminate between the reconstructed image in the generator and the real high-resolution image,improving the model capability in the process of discriminating with the real high-resolution image.At the same time,the effect of adding perceptual loss on super-resolution reconstructed images is investigated,making the reconstructed images more rich in texture details.The experimental results show that the peak signal-to-noise ratio of the reconstructed image is improved by 0.48 dB and the structural similarity is improved by 0.023 compared with the original algorithm.The model is improved in subjective visual evaluation and objective quantitative assessment.

关 键 词：深度学习 超分辨 SRGAN 生成器 判别器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]