If you have any questions, please feel free to contact us.

email me at: jacksung1995@gmail.com

or create a new issue directly.

LGAN

Information

Source codes for "A Lightweight Local-Global Attention Network for Single Image Super-Resolution". It is accepted by ACCV2022.

The structure of LGAN is shown as follows:

The structure of LGAB is shown as follows:

How the feature maps reshaped is described in the figure:

The visual comparison is shown as follows:

Performance Evaluation:

Please check the paper for more details.

Steps

Step 1: Install environment

The required environment is listed in 'env.yaml', please install the environment before training.

Step 2: Edit config file

Settings can be edited in 'configs/config.yml'. If you do not want to change the settings, you can rename the 'configs/config_default.yml' to 'configs/config.yml'.

Step 3: Prepare the dataset

You can download the dataset from here.

|/SR_datasets
|-|/DIV2K
|-|-|/DIV2K_train_HR/
|-|-|/DIV2K_train_LR_bicubic/
|-|/benchmark
|-|-|/B100
|-|-|-|/HR
|-|-|-|/LR_bicubic
|-|-|/Manga109
|-|-|-|/HR
|-|-|-|/LR_bicubic
|-|-|/Set14
|-|-|-|/HR
|-|-|-|/LR_bicubic
|-|-|/Set5
|-|-|-|/HR
|-|-|-|/LR_bicubic
|-|-|/Urban100
|-|-|-|/HR
|-|-|-|/LR_bicubic

Step 4: Train the network

Run

sh train.sh

or you can download pre-trained models x2 x3 x4.

NOTE: PSNR and SSIM listed in paper is calculated by matlab program. The result calculated by Python program may be slightly inconsistent with the paper.

Step 5: Evaluate the network

To evaluate the network, you should specify the parameter 'pretrain' in 'configs/config.yaml' first. And run

sh test.sh

The visual result produced by LGAN can be obtained here.

Thanks

The codes are implemented based on ELAN.