Yue Qiuyu, Chen Bingliang, Liu Xinyu, Zheng Zhenrong
Opt Lett. 2024 Jun 1;49(11):2894-2897. doi: 10.1364/OL.523832.
Coded aperture snapshot spectral imaging (CASSI) can capture hyperspectral images (HSIs) in one shot, but it suffers from optical aberrations that degrade the reconstruction quality. Existing deep learning methods for CASSI reconstruction lose some performance on real data due to aberrations. We propose a method to restore high-resolution HSIs from a low-resolution CASSI measurement. We first generate realistic training data that mimics the optical aberrations of CASSI using a spectral imaging simulation technique. A generative network is then trained on this data to recover HSIs from a blurred and distorted CASSI measurement. Our method adapts to the optical system degradation model and thus improves the reconstruction robustness. Experiments on both simulated and real data indicate that our method significantly enhances the image quality of reconstruction outcomes and can be applied to different CASSI systems.
编码孔径快照光谱成像（CASSI）能够一次性捕获高光谱图像（HSIs），但它存在光学像差，会降低重建质量。现有的用于CASSI重建的深度学习方法由于像差在实际数据上会损失一些性能。我们提出了一种从低分辨率CASSI测量中恢复高分辨率HSIs的方法。我们首先使用光谱成像模拟技术生成模拟CASSI光学像差的逼真训练数据。然后在这些数据上训练一个生成网络，以从模糊和失真的CASSI测量中恢复HSIs。我们的方法适应光学系统退化模型，从而提高了重建的鲁棒性。在模拟数据和实际数据上的实验表明，我们的方法显著提高了重建结果的图像质量，并且可以应用于不同的CASSI系统。
