Zero-order diffraction-guided spectral-wise transformer: a lightweight and efficient framework for computed tomography imaging spectrometry

Computed tomography imaging spectrometry (CTIS) is an advanced snapshot hyperspectral imaging technique capable of acquiring the three-dimensional (x, y, <i>λ</i>) datacube of the target scene within a single exposure. Due to the limited number of projections and constrained projection angles, the reconstructed datacube usually suffers severe artifacts. In recent years, deep learning has been applied in hyperspectral datacube reconstruction and has shown superiority in suppressing artifacts. However, most existing methods are CNN-based and contain a substantial number of parameters, necessitating large datasets for training to avoid overfitting. In this paper, we present a lightweight and efficient framework, Zero-order Diffraction-guided Spectral-wise Transformer (ZDST), which is capable of achieving high reconstruction quality with low computation resource consumption and a limited dataset by integrating three key modules: Full Projection Connect Reshaping (FPCR), Spectral-wise Multi-head Self-Attention (S-MSA), and a Zero-order Diffraction-guided Mechanism (ZDM). These modules effectively incorporate optical calibration information, spectral-wise similarity, and zero-order diffraction information into the network architecture, respectively, which enhances the fitting capability of the model for the CTIS reconstruction task and reduces the model's reliance on extensive parameters. Additionally, we collected a real dataset using a custom-built filter wheel system. Experimental results demonstrate that our model performs well under real conditions.