Stratified-Attention-Dense-Network

This repository is for SADN introduced in the following paper

ZhiWei Liu,XiaoFeng Mao,Ji Huang,MengHan Gan andYueYuan Zhang, "Stratified Attention Dense Network for Image Super-Resolution".

Requirement

python >= 3.7
tensorflow == 1.14
numpy == 1.15.4

Introduction

Stratified Attention Dense Network was proposed to reconstruct high quality HR image. Stratified Attention Dense Network

Attention Dense Module

Training

1.Download DIV2K training data(800 training + 100 validation images) from DIV2K dataset

Begin to train

Example command is in the file 'demo.txt'.

python main.py --train_GT_path F:/ARDN/DataSet/DIV2K_train_HR --train_LR_path F:/ARDN/DataSet/DIV2K_train_LR_bicubic/X2/ --test_GT_path F:/ARDN/DataSet/benchmark/Set5/HR/ --test_LR_path F:/ARDN/DataSet/benchmark/Set5/LR_bicubic/X2/ --test_with_train True --scale 2 --log_freq 1 --model_save_freq 10000 --max_step 10000 --n_ARDG 12 --n_ARDB 12

Test

Quick start

1.Download benchmark datasets from SNU_CVLab

(optional) Download pretrained models for our paper.

The Trained x2 models can be downloaded from Google Drive
Run the following scripts. Example command is in the file 'demo.txt'.

python main.py --mode test --pre_trained_model ./model/ARDN_X2_64_8_8-1000000 --test_GT_path F:/ARDN/DataSet/benchmark/Urban100/HR/ --test_LR_path F:/ARDN/DataSet/benchmark/Urban100/LR_bicubic/X2/ --scale 2 --save_test_result False --test_set Urban100+ --self_ensemble True --chop_forward True

maple-mao/Stratified-Attention-Dense-Network