splinter21 / BAM

we proposed a light-weigt but efficient blanced attention mechanism for SISR task.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

BAM (The project is a bit messy, I will clean up the code later)

This project is built from IDN, and thanks for the contributions of all the other researchers those who made their codes accessible.

realSR7

baidu link:https://pan.baidu.com/s/1UmUc7Gbjip72rvLpDRkk6Q passwords:cxs8

Requirements

  • PyTorch>=1.0.0
  • Numpy 1.15.4
  • Pillow 5.4.1
  • h5py 2.8.0
  • tqdm 4.30.0

result

Some of the indicators in the MSRN paper are quite different from the indicators given in the original author's open source code, and the indicators in the open source code are used:https://github.com/MIVRC/MSRN-PyTorch

Examples

realSR7

Different attention mechanisms comparison

绘图1 绘图3_100

IMDN VS IMDN-BAM

绘图2_100

IMDN VS IMDN-BAM && DRLN VS DRLN-BAM

图片

structure

图片

control exp

图片

图片

图片

visual and metric comparison

图片

visual comparison

图片

metric comparison

图片

ablation exp

图片

图片

speed exp

FPS comparison

图片

Train

The DIV2K, Set5 dataset converted to HDF5 can be downloaded from the links below.Otherwise, you can use prepare.py to create custom dataset.

Dataset Scale Type Link
DIV2K 2 Train Download
DIV2K 3 Train Download
DIV2K 4 Train Download
Set5 2 Eval Download
Set5 3 Eval Download
Set5 4 Eval Download

The Flickr2K dataset can be downloaded from the links below,and then you can use prepare.py to create custom dataset. https://link.csdn.net/?target=http%3A%2F%2Fcv.snu.ac.kr%2Fresearch%2FEDSR%2FFlickr2K.tar

for preparedata: python3 prepare.py --images-dir ../../DIV2K/DIV2K_train_HR --output-path ./h5file_DIV2K_train_HR_x4_train --scale 4 --eval False

for train: python3 train.py --choose_net DRLN_BlancedAttention --train_file ./h5file_mirflickr_train_HR_x3_train --eval_file ./h5file_Set5_x4_test

for eval all SR size && all networks(you should download checkpoints first); python3 eval_allsize_allnet.py

for eval dingle image: python3 eval_singleimg.py --lr_image_file ./savedimg/Set5/4/EDSR_blanced_attention_2.png --hr_image_file ../classical_SR_datasets/Set5/Set5/butterfly.png

for infer all size && all networks SR images(the SR images will be saved in the direct ./savedimg/*): python3 infer_allsize_allnet.py

##checkpoints We provide all network && all size checkpoints to prove that our experiments are convincing. you can get them from:https://pan.baidu.com/s/1gy-3jcikT2h-QfRduwoibg password: 2ubm

If the password fails or any other questions, please contact me:2463908977@qq.com

##Our Attention mechanism is very tiny and efficient, and has also been proved to be efficient in semantic segmentation missions,especially for light-weight models.

ezoic increase your site revenue

About

we proposed a light-weigt but efficient blanced attention mechanism for SISR task.

License:MIT License


Languages

Language:Python 100.0%