基于自适应深度先验的高光谱图像超分辨率

Adaptive deep prior for hyperspectral image super-resolution

为了解决现有的高光谱超分辨率方法依赖于手工先验和数据驱动先验会导致参数选择困难或可解释性差的问题,采用一种基于自适应深度先验正则的高光谱图像超分辨率方法,进行了理论分析和实验验证。首先设计基于卷积神经网络的多阶段特征提取网络,提取退化图像的空间和光谱信息;其次将提取到的空-谱先验输入基于transformer模型的特征融合模块;然后自适应交互空域和谱域的互补信息,以捕获图像的全局先验特征;最后在退化模型中插入深度先验正则项,将超分辨率问题表述为一个优化问题,其解可以通过交替方向乘子法获得并降低求解复杂度。结果表明,所提出算法在信噪比均为35 dB时,重建信噪比分别达到了34.16 dB和29.35 dB,比次优算法高出2.78 dB和2.17 dB,重建的高分辨率高光谱图像与其固有结构具有较高的一致性。该研究为综合利用手工先验和数据驱动先验增强高光谱图像空间分辨率提供了参考。

To address the issue that the difficult parameter selection or poor interpretability caused by the existing hyperspectral super-resolution methods rely on manual or data-driven prior,a deep prior regularization-based hyperspectral image super-resolution method was adopted for theoretical analysis and experimental validation.First,a multi-stage feature extraction network based on a deep convolutional neural network was designed to extract spatial and spectral information from the degraded images.Then,the collected spatial-spectral prior features were fed into a transformer-based feature fusion module,where complementary information from the spatial and spectral domains was adaptively extracted to capture the image’s global prior features.Finally,the super-resolution problem of the image was formulated as an optimization problem by inserting deep prior regularization term in the degraded model,the solution of which can be achieved using the alternating direction method of multipliers while minimizing solution complexity.Experimental results show that reconstruction signal-to-noise ratio of this algorithm is 34.16 dB and 29.35 dB when both of the signal-to-noise ratio are 35 dB,which is 2.78 dB and 2.17 dB higher than the suboptimal algorithm,respectively.The reconstructed high-resolution hyperspectral images have high consistency with their inherent structures under the condition of deep prior regularization.This study provides a reference for the combined use of manual and data-driven prior to enhance the spatial resolution of hyperspectral images.