基于频域-空域混合注意力的降质遥感图像质量增强

Degradation remote sensing image quality enhancement based on frequency-domain-spatial-domain hybrid attention

针对传统图像处理算法稳定性差,而深度学习算法因为缺少用于遥感图像增强的训练数据集且算法计算量大导致无法满足遥感图像处理的工程应用需求。综合考虑退化建模和图像处理两个过程,基于Zernike多项式退化模型提出了一种构建遥感图像增强数据集的方法。然后设计了一种基于频域-空域混合注意力的降质遥感图像质量增强算法,在算法的结构设计中通过引入双域选择模块、频率特征残差模块使得在频域和空域分别增强了对高频信息丰富的图像纹理、细节等特征的学习,并且通过使用混合注意力机制提高了模型的特征提取能力。为了验证本文所提算法的性能,以NIQE数值、可视化效果、MTF曲线、推理时效性作为降质遥感图像质量增强的评价指标,比较了本文算法与五种经常使用的算法分别在高分二号卫星影像数据上的图像质量增强结果。实验结果表明,相比于常用方法,本文方法能够更好地降低NIQE数值和提高MTF曲线,具有更清晰地可视化效果,因此能够显著提升降质图像的质量;此外在推理时效性方面,在像素尺寸为27620×29200的遥感图像的处理时间上,相较于传统算法需要时间以小时计,而本文所提算法处理速度仅为27 s,因此能够满足工程应用的时效性要求。综上所述,本文研究可以实现遥感图像的快速清晰化处理,为在轨卫星成像退化问题提供了有效的解决方案。

Traditional image processing algorithms lack stability,and deep learning algorithms fall short of engineering requirements for remote sensing due to insufficient training datasets and high computational de⁃mands.To tackle this,the paper integrates degradation modeling and image processing.It introduces a method for creating a remote sensing image enhancement dataset using a Zernike polynomial degradation model.Additionally,it designs an algorithm for enhancing degraded remote sensing images using a hybrid frequency-domain and spatial-domain attention mechanism.This algorithm employs a dual-domain selec⁃tion module and a frequency feature residual module to improve the learning of high-frequency image tex⁃tures and details in both domains.The hybrid attention mechanism further boosts feature extraction capa⁃bilities.The algorithm's performance was validated against five common methods using NIQE values,vi⁃sualization effects,MTF curves,and inference efficiency.Results show that the proposed approach signifi⁃cantly reduces NIQE values and improves MTF curves,leading to clearer images and substantially en⁃hancing degraded image quality.For images with a specific pixel size of 27620×29200,the algorithm processes them in just 27 s,compared to hours for traditional methods,thus meeting engineering timeliness re⁃quirements.This research offers a rapid and effective solution for addressing satellite imaging degradation.