Table 2 Comparison of neural network-based works for end-to-end spectral reconstruction from RGB images

From: Spectral imaging with deep learning

Method

Network Backbone

Deep-Learning Techniques

Evaluation (RMSE)

Evaluation (MRAE)

HSCNN23

CNN

Multi-layer CNN

17.006 on Clean track (NTIRE 2018)

0.0190 on Clean track (NTIRE 2018)

Adv_rgb2hs62

cGAN

Additional MAE loss on cGAN

1.457 on ICVL, 24.81 on Clean, 34.05 on Real World (NTIRE 2018)

0.0218 on Clean, 0.0396 on Real World (NTIRE 2018)

Spectral super-resolution60

Densenet

Dense structure, sub-pixel convolution layer

1.98 on ICVL, 5.27 on NUS, 4.76 on CAVE

/

HSCNN+71

2-level CNN

Residual blocks, dense structure, feature fusion70

13.128[14.45] on Clean, 22.935[24.06] on Real World (NTIRE 2018)

0.0135[0.0137] on Clean, 0.0293[0.0310] on Real World (NTIRE 2018)

LFB98

U-Net

Camera sensitivity prior

20.146[16.19] on Clean, 27.557[26.44] on Real World (NTIRE 2018)

0.01704 on Clean, 0.03081 on Real World (NTIRE 2018)

CVL99

2-level CNN

Residual blocks, PReLU activation

1.23 on ICVL, 3.66 on NUS, 3.5275 on CAVE, 17.27 on Clean, 27.09 on Real World (NTIRE 2018)

0.0174[0.0152] on Clean, 0.0364[0.0335] on Real World (NTIRE 2018)

3D-CNN100

CNN

3D CNN101

1.115 on ICVL, 2.86 on CAVE, 20.010[19.41] on Clean (NTIRE 2018)

0.018[0.0181] on Clean track (NTIRE 2018)

Sensitivity estimation79

CNN

Back-projection loss

(s) 0.0282 on ICVL, 0.0316 on CAVE

(s) 0.13 on ICVL 0.38 on CAVE

Deep Function-Mixture Network68

multi-level CNN

Pixel attention, feature fusion

4.54 on CAVE, 2.54 on Harvard, 1.03 on NTIRE 2018, 0.01268 on Clean, 0.01946 on Real World (NTIRE 2020)

0.03075 on Clean, 0.06212 on Real World (NTIRE 2020)

MXR-U-Nets102

U-Net

XResnet block103, Mish activation function104, feature and style loss, self-attention layer

0.01645 on Clean, 0.02255 on Real World (NTIRE 2020)

0.0454[0.04441] on Clean, 0.0840[0.09322] on Real World (NTIRE 2020)

AWAN80

CNN

Residual blocks, channel attention, PReLU activation, back-projection loss, self ensemble, model ensemble

0.0111[0.01293] on Clean, 0.0170[0.01991] on Real World (NTIRE 2020); 10.24 on Clean, 21.33 on Real World (NTIRE 2018)

0.0312[0.03010] on Clean, 0.0639[0.06210] on Real World (NTIRE 2020), 0.0114 on Clean, 0.0277 on Real World (NTIRE 2018)

RPAN72

CNN

Pixel attention, global residual connection, feature fusion

4.301[0.01695] on Clean, 4.984[0.02071] on Real World (NTIRE 2020)

0.03756[0.03601] on Clean, 0.06787[0.06780] on Real World (NTIRE 2020)

HRNet73

4-level CNN

Residual blocks, dense structure, feature fusion, sub-pixel convolution layer, model ensemble

0.01354[0.01389] on Clean, 0.01786[0.01923] on Real World (NTIRE 2020)

0.04233[0.03231] on Clean, 0.06825[0.06200] on Real World (NTIRE 2020)

C2H-Net92

U-Net

Additional category prior

4.7313 on CAVE

/

Double Ghost105

GhostNet

GhostNet block106, residual connection, attention mechanism, non-local block107, PReLU activation

0.0162 on Real World track (NTIRE 2020)

0.0439 on Real World track (NTIRE 2020)

  1. In the network architecture column, level means parallel CNN layers for data flow. For the deep-learning techniques column, we highlight the techniques that may play an important role in the method’s performance. Performance evaluations are collected from reported results of the original article, corresponding articles or NTIRE competition. Evaluation results from the original article is considered first, then NTIRE competition, and finally the corresponding articles. If the evaluation results occurred in both the original article and the NTIRE competition report, we use [] to denote the evaluation result in the NTIRE report. Evaluation values labeled “s” in the table are from the scaled dataset (datasets that are linearly scaled to [0,1] range)
Back to article page