«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年08期

页码:

151-157

栏目:

人工智能

出版日期:

2024-08-10

文章信息/Info

Title:

Research on Super Resolution Reconstruction of Remote Sensing Images Based on Multi-scale Convolutional Residuals

文章编号:

1673-629X(2024)08-0151-07

作者:

张剑飞*; 屠艳杰

黑龙江科技大学 计算机与信息工程学院,黑龙江 哈尔滨 150022

Author(s):

ZHANG Jian-fei*; TU Yan-jie

School of Computer and Information Engineering,Heilongjiang University of Science and Technology,Harbin 150022,China

关键词:

遥感图像; 超分辨率; 生成对抗网络; 自适应归一化; 注意力机制

Keywords:

remote sensing images; super resolution; generative adversarial networks; adaptive normalization; attention mechanism

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0131

摘要:

针对现有遥感图像超分辨率重建技术在纹理细节恢复、伪影抑制和模型收敛等方面存在的问题,提出一种基于生成对抗网络的超分模型 RS-SRGAN(Residual-SN-SimAM SRGAN)。 首先,在生成器中应用多尺度卷积的密集剩余残差卷积块(Residual-in-Residual Density Conv Block,RRDCB) 进行深层特征提取,更好地恢复图像细节,同时去除批归一化(BN)提高模型的泛化能力。 然后,在判别器中引入自适应归一化(SN)层替代传统的 BN 层,使网络能够自适应地提取图像特征,并加速模型的收敛。 最后,通过集成无参的 SimAM 注意力机制,增强判别器捕获和理解图像中的关键局部细节能力,在不额外增加参数的基础上有效提高模型的鉴别能力,进一步提高图像的生成质量。 实验结果表明,与原始的 SRGAN相比,改进的模型在 UCMLU 和 NWPU 数据集上的峰值信噪比(PSNR)分别提升了 1. 075 4 dB 和 0. 349 2 dB,结构相似性 (SSIM)分别提升了 0. 004 9 和 0. 007 0。 该研究为遥感图像超分辨率的研究和应用提供了新的视角和技术基础。

Abstract:

Aiming at the problems of existing remote sensing image super resolution ( SR) reconstruction techniques in texture detail recovery,artifact suppression, and model convergence, an improved model based on a generative adversarial network, RS - SRGAN (Residual-SN-SimAM SRGAN) is proposed. Firstly,a dense residual convolution block (Residual-in-Residual Density Conv Block,RRDCB) with multi-scale convolution is applied to the generator for deep feature extraction to better recover the image details,while re-moving the batch normalization (BN) to improve the generalization ability of the model. Then,an adaptive normalization (SN) layer is introduced into the discriminator to replace the traditional BN layer, enabling the network to adaptively extract image features and accelerating the convergence of the model. Finally,by integrating the parameter-free SimAM attention mechanism,the discriminator is enhanced to capture and understand the key local details in the image,which effectively improves the discriminative ability of the model without additional parameters and further improves the quality of image generation. The experimental results show that compared with the original SRGAN,the improved model improves the peak signal-to-noise ratio (PSNR) by 1. 075 4 dB and 0. 349 2 dB on the UCMLU and NWPU datasets,respectively,and the structural similarity (SSIM) by 0. 004 9 and 0. 007 0,respectively. This study provides new research and application of the super-resolution of remote sensing images with a new perspective and technical basis.

常用功能

更新日期/Last Update: 2024-08-10