Author Affiliations
1Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190, China2University of Chinese Academy of Sciences, Beijing 100049, China3Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA4Beijing University of Posts and Telecommunications, Beijing 100876, China5New Jersey Institute of Technology, Newark, New Jersey 07102, USA6Key Laboratory for Quantum Optics and Center for Cold Atom Physics of CAS, Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, Shanghai 201800, China7Center of Materials Science and Optoelectronics Engineering, University of Chinese Academy of Sciences, Beijing 100049, China8Nokia Bell Labs, Murray Hill, New Jersey 07974, USAshow less
Fig. 1. Generalized image formation (left) and the discrete matrix-form model (right) of spectral SCI. Here color denotes the corresponding spectral band.
Fig. 2. Comparison of image-plane coding (upper) and aperture-plane coding (lower) spectral SCI systems in terms of sensing matrix. Here each color block denotes the corresponding transport matrix at that spectral band.
Fig. 3. Image formation process of a typical spectral SCI system, i.e., SD-CASSI and the reconstruction process using the proposed deep PnP prior algorithm.
Fig. 4. Network structure of the deep spectral denoising prior.
Fig. 5. Test spectral data from (a) ICVL [
69] and (b) KAIST [
35] data sets used in simulation. The reference RGB images with pixel resolution
256×256 are shown here. We crop similar regions of the whole image for spatial sizes of
512×512 and
1024×1024.
Fig. 6. Simulation results of color-checker with size of 256×256 from KAIST data set compared with the ground truth. PSNR and SSIM results are also shown for each algorithm.
Fig. 7. Simulation results of exemplar scenes (top, ICVL; bottom, KAIST) with size of 256×256 compared with the ground truth. Spectral curves of selected regions are also plotted to compare with the ground truth.
Fig. 8. Simulation results of four selected scenes shown in sRGB and spectral curves with spatial size of 512×512 (shown in full size in the far left column). The spectra of the pinned (yellow) region of the close-up are shown on the right.
Fig. 9. Simulation results of four selected scenes shown in sRGB and spectral curves with spatial size of 1024×1024 (shown in full size in the far left column). The spectra of the pinned (yellow) region of the close-up are shown on the right.
Fig. 10. Real data, object SD-CASSI data (256×210×33).
Fig. 11. Real data, bird SD-CASSI data (1021×731×33).
Fig. 12. Real data, Lego SD-CASSI data (660×550×28).
Fig. 13. Real data, plant SD-CASSI data (660×550×28).
Fig. 14. Real data, snapshot multispectral endomicroscopy data (660×660×24).
Fig. 15. Real data, GISC spectral camera data (330×330×16).
Spatial Size | Data Set | TwIST | GAP-TV | AE | U-net | PnP | PSNR (dB) | SSIM | Running Time (s) | PSNR (dB) | SSIM | Running Time (s) | PSNR (dB) | SSIM | Running Time (s) | PSNR (dB) | SSIM | Running Time (s) | PSNR (dB) | SSIM | Running Time (s) | | ICVL | 30.58 | 0.8731 | 156.3 | 32.57 | 0.8794 | 130.2 | 29.41 | 0.8711 | 144.2 | 31.13 | 0.8897 | 0.8 | 35.03 | 0.9274 | 132.7 | KAIST | 27.32 | 0.8495 | 29.66 | 0.8584 | 26.79 | 0.8498 | 29.44 | 0.8941 | 33.21 | 0.9273 | | ICVL | 31.82 | 0.8955 | 1380.2 | 33.58 | 0.8965 | 399.1 | 31.22 | 0.8969 | 493.6 | NA | NA | NA | 35.68 | 0.9319 | 401.6 | KAIST | 29.09 | 0.8944 | 31.38 | 0.8993 | 29.28 | 0.8974 | NA | NA | 34.29 | 0.9378 | | ICVL | 32.68 | 0.9159 | 3657.6 | 34.22 | 0.9157 | 1460.7 | 32.03 | 0.9158 | 2053.5 | NA | NA | NA | 36.21 | 0.9434 | 1453.6 | KAIST | 31.64 | 0.9099 | 33.66 | 0.9134 | 31.05 | 0.9071 | NA | NA | 36.41 | 0.9433 |
|
Table 1. Average PSNR (in dB), SSIM, and Running Time (in Seconds) of 16 Simulation Scenes (8 from ICVL and 8 from KAIST) at Different Spatial Sizes Using Various Algorithmsa