Multiscale-Super-Spectral

---

Bachelor's Thesis
English Version
Chinese Version (Unofficial)

Project webpabe
Spectral Super-resolution from Single RGB Image Using Multi-scale CNN

---

Pre-trained models

Download pre-trained model from Baidu Cloud or Google Drive
Place them at TestLog.

Contents of this repository

• dataset.py: data pre-processing
• model.py & train.py: our model
• model_ref.py & train_ref.py: the comparison model
• test.py: test the models and report the evaluation results
• utilities.py: some auxiliary functions

How to run

1. Dependences

• PyTorch (<0.4)
• OpenCV for Python3
• HDF5 for Python
• tensorboardX
• Matplotlib

2. Download data

NTIRE 2018 challenge on spectral reconstruction from RGB images (Track 1)

Place the data at data and arrange the directories as follows:

data/Train_Spectral/
--BGU_HS_00001.mat
... ...
--BGU_HS_00256.mat

data/Train_RGB/
--BGU_HS_00001_clean.png
... ...
--BGU_HS_00256_clean.png

data/Test_Spectral/
--BGU_HS_00257.mat
--BGU_HS_00259.mat
--BGU_HS_00261.mat
--BGU_HS_00263.mat
--BGU_HS_00265.mat

data/Test_RGB/
--BGU_HS_00257_clean.png
--BGU_HS_00259_clean.png
--BGU_HS_00261_clean.png
--BGU_HS_00263_clean.png
--BGU_HS_00265_clean.png

3. Train our model

If you are running the code for the first time, remember data pre-processing.

python3 train.py --preprocess True

Otherwise, run training directly.

python3 train.py

You may also adjust hyper-parameters such as batch size, initial learning rate, dropout rate, etc.

4. Train the comparison model

Reference paper: Learned Spectral Super-resolution
The data pre-processing is exactly the same as above, so you can run training directly.

python3 train_ref.py

5. Test our model

Dropout rate: 0.2

python3 test.py --model Model \
  --dropout 2

Dropout rate: 0

python3 test.py --model Model \
  --dropout 0

6. Test the comparison model

Dropout rate: 0.5

python3 test.py --model Ref \
  --dropout 5

Dropout rate: 0

python3 test.py --model Ref \
  --dropout 0