Fast-super-resolution-by-CNN

Reimplementation of https://arxiv.org/abs/1608.00367 in Tensorflow 2.1.

Authors

• Max Holmberg
• Joel Lidin
• Samuel Norling

Changes

• Used Adam as the optimizer instead of SGD

Prepare the data

For the 91-image dataset

python extract_patches.py -path "dataset/T91" -output_path "T91_x3.h5" -f_sub_lr 7 -upscaling 3

For the validation dataset (20 random images from BSD500)

python extract_patches.py -path "dataset/BSD500_val_20" -output_path "BSD500_x3.h5" -f_sub_lr 7 -upscaling 3

Training

To train from scratch

python run.py -epochs 15 [-include_test] -train_path "T91_x3.h5" -val_path "BSD500_x3.h5" -f_sub_lr 7 -upscaling 3 -batch_size 128

To resume training (pretrained weights for upscaling factor of 3 and 4 are included in weights_x3 and weights_x4)

python run.py -epochs 15 -continue -weights weights_x3 [-include_test] -train_path "T91_x3.h5" -val_path "BSD500_x3.h5" -f_sub_lr 7 -upscaling 3 -batch_size 128 

Results

Test dataset	upscaling factor	bicubic	FSRCNN (Dong et al.)	FSRCNN (Our with Adam)
Set5	3	30.91	33.06	33.79
Set14	3	27.91	29.37	29.85
BSD200	3	27.70	28.55	29.14
Set5	4	28.47	30.55	29.75
Set14	4	25.99	27.50	26.91
BSD200	4	26.66	26.92	27.37

Images

Original	Bicubic	FSRCNN (with Adam)

Original	Bicubic	FSRCNN (with Adam)