Table 1 Comparison of different amplitude-coded compressive spectral imaging methods
Article | CSI architecture | Performance (PSNR) | Reconstruction model | Deep-learning techniques |
|---|---|---|---|---|
AutoEncoder30 | SD/DD/SS CASSI | 32.46 on CAVE (SS-CASSI) | Autoencoder Equation (Eq. (11) in ref. 30) | Autoencoder prior |
HyperReconNet39 | SD CASSI | 33.63 on ICVL, 31.36 on Harvard | CNN | Hardware representation layer (joint training) |
Spatial–spectral prior24 | SD CASSI | 34.13 on ICVL, 32.84 on Harvard, 30.03 on KAIST | Unrolled network | Learned network prior |
External–internal learning35 | SD CASSI | 35.884 on ICVL, 33.585 on Harvard, 29.055 on CAVE | CNN | Dense structure, back-projection pixel loss |
λ-Net36 | SD CASSI | 32.29 on ICVL (average of 16 scenes) | conditional GAN | Self-attention, hierarchical structure |
DNU42 | SD CASSI | 34.24 on ICVL, 32.71 on Harvard | Unrolled network | Learned network prior |
HCS2-Net31 | SD/SS CASSI | 34.52 on ICVL (10 scenes), 39.22 on CAVE (SS-CASSI), 29.33 on CAVE (SD-CASSI) | CNN (untrained) | Residual block, attention module, unsupervised learning, hardware code concatenated to the input measurement, deep image prior |
Deep-Tensor45 | SD CASSI | 30.92 on ICVL, Harvard and KAIST (best mean) | CNN (untrained) | Learned tensor decomposition |
