首页|轻量级重参数化的遥感图像超分辨率重建网络设计

轻量级重参数化的遥感图像超分辨率重建网络设计

扫码查看
原文链接
  • 万方数据
  • 维普
针对当前基于深度学习的遥感图像超分辨率重建模型部署时对硬件要求较高,本文设计了一种轻量级基于重参数化的残差特征遥感图像超分辨率重建网络.首先,采用重参数化方法设计了一种残差局部特征模块,以有效地提取图像局部特征;同时考虑到图像内部出现的相似特征,设计了一个轻量级的全局上下文模块对图像的相似特征进行关联以提升网络的特征表达能力,并通过调整该模块的通道压缩倍数来减少模型的参数量和改善模型的性能;最后,在上采样模块前使用多层特征融合模块聚合所有的深度特征,以产生更全面的特征表示.在UC Merced遥感数据集上进行测试,该算法在遥感图像3倍超分辨率下的参数量为539 K,峰值信噪比为30.01 dB,结构相似性为0.844 9,模型的推理时间为 0.010 s;而HSENet算法的参数量为 5 470 K,峰值信噪比为 30.00 dB,结构相似性为 0.842 0,模型的推理时间为0.059 s.实验结果表明,该算法相比HSENet算法,参数量更少,运行速度较快,且峰值信噪比与结构相似性也有一定的提高.在DIV2K自然图像数据集上进行测试,该算法的峰值信噪比和结构相似性相比其他算法也有一定的优势,表明该算法的泛化能力较强.
Design of lightweight re-parameterized remote sensing image super-resolution network
In response to the high hardware requirements associated with the deployment of current deep learning-based remote sensing image super-resolution reconstruction models,this paper presented a light-weight,re-parameterized residual feature remote sensing image super-resolution reconstruction network.Firstly,a residual local feature module was designed using re-parameterization to effectively extract local image features.Simultaneously considering the occurrence of similar features within images,a lightweight global context module was devised to associate similar features in images,enhancing the network's feature representation capability.The channel compression rate of this module was adjusted to reduce the model's parameter count and improve its performance.Finally,a multi-level feature fusion module was employed before the upsampling module to aggregate deep features and generate a more comprehensive feature repre-sentation.Tested on the UC Merced remote sensing dataset,this algorithm exhibits a parameter count of 539 K for×3 super-resolution,a PSNR of 30.01 dB,a SSIM of 0.844 9,and an inference time of 0.010 s.In comparison,the HSENet algorithm has a parameter count of 5 470 K,a PSNR of 30.00 dB,an SSIM of 0.842 0,and an inference time of 0.059 s.Experimental results demonstrate that this algo-rithm outperforms the HSENet algorithm,featuring fewer parameters,faster execution,and notable im-provements in PSNR and SSIM.Testing on the DIV2K natural image dataset reveals that this algorithm exhibits advantages in PSNR and SSIM compared to other algorithms,demonstrating its strong generaliza-tion capability.

super resolutionremote sensing imagesglobal contextre-parameterizationresidual net-work

易见兵、陈俊宽、曹锋、李俊、谢唯嘉

展开 >

江西理工大学 信息工程学院,江西 赣州 341000

超分辨率 遥感图像 全局上下文 重参数化 残差网络

国家自然科学基金资助项目国家自然科学基金资助项目江西省自然科学基金资助项目江西省教育厅科技项目资助江西省教育厅科技项目资助江西省教育厅科技项目资助江西省赣州市科技计划资助项目江西省研究生创新专项资助

620660187226101820181BA B202004GJJ210828GJJ200818GJJ180482YC2022-S640

2024

10.37188/OPE.20243202.0268

光学精密工程
中国科学院长春光学精密机械与物理研究所 中国仪器仪表学会

光学精密工程

CSTPCD北大核心
影响因子：2.059
ISSN：1004-924X
年,卷(期)：2024.32(2)
  • 1
  • 8