BCM Framework for Image Deblurring

Please download GoPro dataset into Desktop and modify the path in modifier.py and run https://drive.google.com/file/d/1H0PIXvJH4c40pk7ou6nAwoxuR4Qh_Sa2/view

cd datas
python modifier.py

Requires.

pytorch (>= 0.4.1)
numpy
scipy
scikit-image
opencv

For setting config, check the tempplate inside expConfig, you can set the custom deblur net here.

# experiment Name / Method
experiment_name: Patch_Deblur_FPN_VAE_PAC_content_condiser_c16

# Some setting about training
isTrain : True
train_epoch : 100
start_train_epoch : 0
batch_size : 6
image_size : 256
gpu : 0
learning_rate : 0.0001

# deblur part
# deblur net (including size) setting
G : 'Patch_Deblur_FPN_VAE_PAC'
channel : 16

# reblur part
# per-pixel conv size & recurrent time
per_pix_kernel: 3
Recurrent_times: 3

For model training with specific config, run following commands.

python main.py

For model testing, change the config with isTrain to False, then run following commands.

python main.py

Follwing link is the implementationWe of CVPR 2021 paper MPRNet with such framework and reach 32.91 in PSNR
https://github.com/hhcs9527/BRMv2-MPRNet