ESPCN with Perceptual Loss

This repository implements the combination of two works:

• Sub-pixel Convolutional Neural Network (ESPCN)
• Perceptual Loss
• Flow Chart of the SR System:

Requirements

• PyTorch 1.10
• Numpy 1.16.1
• Pillow 6.0.0
• h5py 2.9.0
• matplotlib 3.1.1

Datasets

The DIV2K, Set5 dataset converted to HDF5 can be downloaded from the following link: Datasets

The custom dataset can also be created by following the instruction in the link above.

Train

python train.py --training_set ".../DIV2K_x4.h5" \
--val_set ".../Set5_x4.h5" \
--scale 4 \
--patch_size 56 \
--epoch 200

Test

python test.py --weights "best_path.pth" \
--test_img "img.png" \
--scale 4

Results

• 'comic' image results and FFT-Error plots: