Abstract: Traditional Hyper-Spectral Image (HSI) spatio-spectral fusion algorithms usually use static spectral dictionary, in which the dictionary learning and the image fusion are two separate processes, giving rather poor performance when processing noisy spatio-spectral fusion tasks. To address this problem, a new noisy hyper-spectral image spatio-spectral fusion algorithm based on dynamic dictionary learning was proposed，which adopted an iterative strategy that dynamically updates dictionary atoms during the fusion process, collaborating to complete the spatio-spectral fusion and noise removal tasks. Firstly, a coarse denoising process was performed on the input hyper-spectral image and the denoising result was subsequently utilized to initialize the spectral dictionary; Secondly, the sparse coding technique was employed to fuse the two input images with the initialized dictionary, resulting an intermediate fusion image; Then, the intermediate fusion image was fed back to the dictionary learning module to continuously update the dictionary atoms, forming a dynamic spectral dictionary; Finally, by iterating the above process, the final output image was obtained. Experimental results on three remote sensing hyper-spectral image datasets show that the proposed algorithm can effectively improve the spatial resolution of the simulated noisy HSI. At the same time, experimental results on real noisy image bands also indicate that the proposed approach can significantly improve the visual quality of the fused images. For the Cuprite Mine dataset, the peak signal-to-noise ratio (PSNR) has been increased by 8.79dB and 3.47dB respectively compared to the GTNN（Generalized Tensor Nuclear Norm） and AL-NSSR methods, with the Gaussian noise variance being 0.15 and the amplification factor being 8.

Key words: Hyper-Spectral Image (HSI), spatio-spectral fusion, noise, spectral dictionary learning, iterative sparse representation

摘要： 针对传统高光谱图像（HSI）空谱融合算法通常采用静态光谱字典，字典学习与图像融合过程相分离，对含有噪声的空谱融合任务处理效果不佳的问题，提出一种基于动态字典学习的含噪高光谱图像空谱融合算法。该算法采用迭代思想，在融合过程中动态更新字典原子，协作完成空谱融合及噪声去除任务。首先，对输入高光谱图像进行粗去噪，并利用去噪结果初始化光谱字典；其次，利用上述初始化字典对两幅待融合图像进行稀疏表示，得到中间融合结果；再次，将中间融合结果反馈给字典学习模块，不断更新字典原子，构造动态光谱字典；最后，通过迭代以上过程得到最终的输出图像。在3组遥感高光谱图像上的仿真实验结果表明，所提算法能够在提升图像空间分辨率的同时有效去除噪声；同时，在真实含噪图像波段上的实验结果也表明所提算法能够有效提高融合图像的视觉质量。在Cuprite Mine数据集上，高斯噪声方差为0.15且放大倍数为8时，与基于广义张量核范数的方法（Generalized Tensor Nuclear Norm, GTNN）、AL-NSSR方法相比，峰值信噪比（PSNR）分别提升了8.79dB和3.47dB。

关键词: 高光谱图像, 空谱融合, 噪声, 光谱字典学习, 迭代稀疏表示