zbxzc35 / simpledeblurnet

Simple DeblurNet

Description

This our modification on Scale-Recurrent Net for Deblurring. We name it Simple DeblurNet. The following modifications are made on SRN:

• Substitute the conv-LSTM block with a normal conv layer of the same size.

• Abandon the multi-scale recurrent training process in original SRN.

• Add an overall skip connection: directly add the initial input to the output of the last layer, as what it does in DeblurGAN.

• Devise a new Loss function to fit our compact model and training.

The following figure shows the our model's architecture:

A brief visual comparison among different neural deblurring models:

Some metrics:

1. GoPro dataset:

PSNR	WSNR	SSIM	MSSSIM	IFC	NQM	UIQI	VIF	BLIINDS2	BRISQUE	CORNIA	DIIVINE	NIQE	SSE
23.30	23.55	0.86	0.93	2.59	19.84	0.70	0.79	39.73	114.74	122.84	43.28	20.08	44.96

2. Lai et al's:

metrics	real	uniform	non-uniform
PSNR	-	14.17	14.29
SSIM	-	0.39	0.43
IFC	-	0.18	0.34
NQM	-	4.19	5.58
UIQI	-	0.10	0.17
VIF	-	0.10	0.22
BLIINDS2	37.12	45.40	36.30
BRISQUE	116.15	115.39	110.46
CORNIA	123.82	123.97	122.78
DIIVINE	41.52	44.64	40.49
NIQE	20.13	19.71	19.18
SSEQ	43.54	43.61	34.55

Codes：

This implementation is based on Pytorch0.4 with Cuda backend (CPU version is not considered.)

For visulization, we use a third-party package TensorBoardX.

For other package losing, you can easily fix by code debugging.

Some deblurred results:

GoPro + Lai et al., 2016 : https://jbox.sjtu.edu.cn/l/r1Kk9E

How to Run:

To Train:

For training the model, you need to specify the training dataset. Here we use GoPro dataset. Unzip this dataset and set the dataset_dir attribute in train_config.py to the root directory of the dataset.

You can also set other attributes in train_config.py

To run the code, simply type:

python train.py

To Test:

You must first specify the directory where model checkpoint file is saved, e.g. checkpoint is saved in ./. Our testing code only suports for GoPro testing set. For other purpose, you need manually modify the code.

To run the code, type:

python test.py ./