Spectral imaging is a multidimensional information acquisition technology that combines traditional imaging with spectral analysis. Traditional spectral imaging technologies are often complex and costly, making them difficult to popularize in dynamic or transient scenes. In contrast, snapshot spectral imaging technology can capture spatial and spectral data within a single integration cycle of the imaging system. With the development of micro-nano optics, diffractive optical elements (DOEs) have been applied to snapshot spectral imaging due to their small size and high design flexibility, further reducing device volume and hardware costs. However, existing snapshot spectral imaging technologies based on DOEs are susceptible to the effects of diffractive lens fabrication accuracy and various errors during imaging. Moreover, they require sophisticated spectral image reconstruction algorithms, limiting their widespread application under practical conditions. To fully utilize the advantages of diffraction lens, we conduct in-depth research on their imaging and designing principles, error analysis, image acquisition, reconstruction, and deep learning algorithms. A new type of snapshot differentiable coded spectral imaging system is proposed, which can optimize the design of optical systems and achieve high-quality reconstruction of spectral images. The system demonstrates promising results in simulation and practical image restoration, showcasing its practical value.

We introduce a novel approach using a hybrid diffractive-refractive lens scheme, which effectively reduces the microstructure density of DOEs (Fig. 1). This not only shortens the system focal length and decreases DOE fabrication complexity but also enhances the imaging signal-to-noise ratio. Furthermore, it employs a deep unfolding framework alongside an improved Transformer model (DUF-DST, Fig. 2) to facilitate the reconstruction of diffraction spectral images. Building upon this framework, we conduct a comprehensive analysis of error sources in snapshot diffraction spectral imaging systems. This includes fabrication errors during DOE preparation (Figs. 3 to 4), component assembly discrepancies during imaging (Figs. 5 to 6), as well as sensor and environmental noise factors (Fig. 7). Through rigorous quantitative validation experiments, we quantify the magnitude of each error and assess their impact on imaging and final reconstruction outcomes via meticulous modeling and simulation. Finally, throughout the DOE design and reconstruction model training phases, we employ a joint optimization method to effectively mitigate these error sources.

To validate the effectiveness of the aforementioned optical model and spectral image reconstruction method, we conduct simulation tests by establishing a comprehensive image degradation model and reconstruction network framework based on actual experimental parameters. The DOE utilized in this paper is devised using an end-to-end joint optimization method (Fig. 8), which takes partial machining errors into consideration during the design optimization process. Through degradation-reconstruction testing on 30 scenes, the reconstructed results achieve an average peak signal-to-noise ratio (PSNR) of 37.16, a structural similarity index (SSIM) of 0.9881, and a spectral angle mapper (SAM) of 0.0591 (Fig. 9). Comparison with results from four other mainstream image reconstruction models demonstrates that the DUF-DST model employed here exhibits superior reconstruction performance (Fig. 10). Furthermore, to verify the effectiveness of the error suppression method proposed in this paper, a series of indoor and outdoor experiments are conducted (Figs. 12 to 18). These experimental scenarios closely resemble real-world application environments and encompass various analyzed errors and noise. Reconstruction of the original images captured is performed using a reconstruction network optimized based on error considerations. Experimental results indicate that the reconstruction model employed in this paper achieves high-quality restoration of spectral images, and the proposed error suppression method significantly enhances the robustness of the reconstruction algorithm to errors and noise in actual imaging processes.

Addressing the inadequate consideration of errors in the imaging process by current diffractive spectral imaging technology, which leads to limited imaging effects, we introduce a snapshot diffractive spectral imaging system along with a hybrid error suppression method. It systematically examines errors (height map error and graphic structure location error) arising from diffraction lens fabrication and component assembly, as well as system and environmental noise. Based on these error terms, a diffraction degradation model is constructed, and a deep unfolding network is used to reconstruct the diffraction-blurred images. By jointly training the degradation model and reconstruction network, the reconstruction algorithm’s generalization ability to errors and noise is significantly enhanced. Relevant simulations and indoor/outdoor experiments demonstrate that the imaging model, with error suppression and the proposed reconstruction algorithm, effectively enhances the imaging quality of the system in practical application scenarios, achieving high-quality reconstruction of spectral images within a single integration cycle.