Trying to enhance the undertext in the SGP dataset

python 3.7.4

• network building:
  torch 1.4.0
  
• data structures:
  pandas 0.25.1
  numpy 1.17.2
  
• evaluation metrics:
  sklearn 0.23.2
  
• images operation & curves drawing:
  skimage 0.16.2
  matlpotlib 3.1.1
  torchvision 0.5.0
  

• networks/models.py: classes of all networks.
  
• networks/xxx_classify.py: training of a network (xxx indicates the type of network)
  
• networks/xxx_classify_test_roi.py: testing of a network, outputs enhancement reconstruction of a test image (NOTE: please run training before testing)
  
• training data can be put under networks/data/sgp/xxx.csv (for pixel data) and networks/data/sgp/{folio_id}/cropped_roi/ (for cropped image patches)
• intermediate folders created during training: networks/training_log/, networks/model/, networks/reconstructed_xxx/

Images of 024r_029v
Original version, LDA version, AE-enhanced version, 1DConvNet, 2DFConvNet, 3DConvNet-hyb, Conv-hybrid:

Images of 102v_107r
Original version, LDA version, AE-enhanced version, 1DConvNet, 2DFConvNet, 3DConvNet-hyb, Conv-hybrid:

Images of 214v_221r
Original version, LDA version, AE-enhanced version, 1DConvNet, 2DFConvNet, 3DConvNet-hyb, Conv-hybrid:

Stacked Autoencoder [1]

1DConvNet [2]

2DFConvNet

2DConvNet-hyb

Hybrid Convnet [3]

[1] C. Xing, L. Ma, and X. Yang. Stacked denoise autoencoder based feature extraction and classification for hyperspectral images. Journal of Sensors, 2016:e3632943, 2015.

[2] Hu, W., Huang, Y., Wei, L., Zhang, F. and Li, H., 2015. Deep convolutional neural networks for hyperspectral image classification. Journal of Sensors, 2015.

[3] Lee, H. and Kwon, H., 2016, July. Contextual deep CNN based hyperspectral classification. In 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) (pp. 3322-3325). IEEE.