【摘要】 在编码孔径快照光谱成像（CASSI）重构领域，重构质量的提升受到图像先验正则化和优化方法的双重影响。针对传统正则项设计复杂、调参效率低下及现有深度先验正则项感受野局限、难以捕捉长距离依赖等问题，以及基于模型的优化方法重建精度不足、端到端优化方法缺乏解释性的挑战，文章提出了一种创新的基于非局部深度先验的CASSI重构算法。该算法通过构建非局部深度先验网络，在先验正则化方面实现高光谱图像空间与光谱维度的全局信息和长程依赖的深入捕捉；在优化方法上，引入了深度展开网络的多阶段架构，加快了收敛速度，并显著提高了重建质量。仿真结果显示，新算法重建图像的峰值信噪比（PSNR）较传统方法提高了13.12 dB，较最新的深度学习方法提升了1.06 dB；同时，计算复杂度降低了43.4%，参数量减少了53.2%，有效保证了深度学习方法的计算效率。试验验证结果显示新算法能够还原更多细节、恢复更多的结构和纹理，视觉效果较好，进一步验证了该算法的有效性。这一研究可为深度学习在计算光谱成像重构领域的应用提供一定参考。更多还原

【Abstract】 In the field of compressive aperture snapshot spectral imaging(CASSI) reconstruction, the enhancement of reconstruction quality is dually influenced by image prior regularization and optimization methods. Addressing the complexities and inefficiencies in tuning traditional regularizers, as well as the limited receptive fields and difficulties in capturing long-range dependencies of existing deep prior regularizers, and the challenges of insufficient reconstruction accuracy in model-based optimization methods and lack of interpretability in end-to-end optimization approaches, this paper proposes an innovative CASSI reconstruction algorithm based on non-local deep prior. This algorithm, through the construction of a non-local deep prior network, achieves comprehensive capture of global information and long-range dependencies in both spatial and spectral dimensions for hyperspectral images in the aspect of prior regularization. In terms of optimization method, the introduction of multi-stage architecture in the deep unfolding network accelerates convergence and significantly enhances the reconstruction quality. Simulation results demonstrate that the proposed algorithm achieves a peak signal-to-noise Ratio(PSNR) improvement of 13.12 dB over traditional methods and 1.06 dB over the latest deep learning approaches. Concurrently, it reduces computational complexity by 43.4% and the number of parameters by 53.2%, effectively ensuring the computational efficiency of the deep learning method.Experimental validation results indicate that the new algorithm can restore more details, recover more structures and textures, and provide better visual effects, further verifying its effectiveness. The above research offers valuable references for the application of deep learning in computational spectral imaging reconstruction.更多还原