Computational hyperspectral imaging (CHI) is a cutting-edge technique, which plays a pivotal role in breaking through the quality bottleneck of hyperspectral images (HSI). Among the techniques employed in this domain, the coded aperture snapshot spectral imaging (CASSI) system holds widespread recognition. Nevertheless, the imaging capability of CASSI remains limited due to the hardware conditions and the fragility of outcomes associated with the ill-posed blind reconstruction process. To this end, we propose a novel cascaded transformer architecture, termed , specifically crafted for fusion-aware CHI by means of a dual-imaging mechanism. CasFormer facilitates the effective enhancement of hyperspectral imaging quality by fusing RGB images, with a focus on spatial and spectral domains. As the name suggests, CasFormer is primarily composed of a series of cascade-attention blocks, enabling the fusion of high-spatial-resolution RGB images through spatial coherence alignment and the recovery of spectrally sequential information more compactly and accurately. Furthermore, CasFormer incorporates physical constraints through a decoupling-based loss function, ensuring spatial consistency and spectral fidelity in the fusion-aware CHI process. Extensive experiments conducted across multiple datasets demonstrate the superiority of CasFormer in achieving high-quality imaging results compared to SOTA CHI algorithms. Our code and benchmark datasets will be openly accessible at .

中文翻译：

---

CasFormer：用于融合感知计算高光谱成像的级联变压器

计算高光谱成像（CHI）是一项前沿技术，对于突破高光谱图像（HSI）的质量瓶颈发挥着关键作用。在该领域采用的技术中，编码孔径快照光谱成像（CASSI）系统得到了广泛的认可。然而，由于硬件条件以及与不适定盲重建过程相关的结果的脆弱性，CASSI 的成像能力仍然有限。为此，我们提出了一种新颖的级联变压器架构，称为 ，专门通过双成像机制为融合感知的 CHI 设计。 CasFormer 通过融合 RGB 图像，重点关注空间和光谱域，有助于有效增强高光谱成像质量。顾名思义，CasFormer 主要由一系列级联注意力块组成，通过空间相干对齐实现高空间分辨率 RGB 图像的融合，并更紧凑、更准确地恢复光谱序列信息。此外，CasFormer 通过基于解耦的损失函数合并物理约束，确保融合感知 CHI 过程中的空间一致性和光谱保真度。在多个数据集上进行的大量实验证明，与 SOTA CHI 算法相比，CasFormer 在实现高质量成像结果方面具有优越性。我们的代码和基准数据集将在 上公开访问。