/ Deep Learning

A PyTorch implementation of ESPCN based on CVPR 2016 paper

A PyTorch implementation of ESPCN based on CVPR 2016 paper

ESPCN

A PyTorch implementation of ESPCN based on CVPR 2016 paper "Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network".

Requirements

conda install pytorch torchvision -c soumith
conda install pytorch torchvision cuda80 -c soumith # install it if you have installed cuda
  • PyTorchNet
pip install git+https://github.com/pytorch/[email protected]
  • opencv
conda install opencv

Datasets

Train、Val Dataset

The train and val datasets are sampled from VOC2012.
Train dataset has 16700 images and Val dataset has 425 images.
Download the datasets from here(access code:5tzp),
and then extract it into data directory. Finally run

python data_utils.py

optional arguments:
--upscale_factor      super resolution upscale factor [default value is 3]

to generate train and val datasets from VOC2012 with given upscale factors(options: 2、3、4、8).

Test Image Dataset

The test image dataset are sampled from
| Set 5 | Bevilacqua et al. BMVC 2012
| Set 14 | Zeyde et al. LNCS 2010
| BSD 100 | Martin et al. ICCV 2001
| Sun-Hays 80 | Sun and Hays ICCP 2012
| Urban 100 | Huang et al. CVPR 2015.
Download the image dataset from here(access code:xwhy), and then extract it into data directory.

Test Video Dataset

The test dataset are sampled from Jay Chou's Music Videos. Download the video dataset from
here(access code:6rad), and then extract it into data/test/SRF_xx/video
directory, which xx means the upscale factor.

Usage

Train

python -m visdom.server & python train.py

optional arguments:
--upscale_factor      super resolution upscale factor [default value is 3]
--num_epochs          super resolution epochs number [default value is 100]

Visdom now can be accessed by going to 127.0.0.1:8097 in your browser,
or your own host address if specified.

If the above does not work, try using an SSH tunnel to your server by
adding the following line to your local ~/.ssh/config :
LocalForward 127.0.0.1:8097 127.0.0.1:8097.

Maybe if you are in China, you should download the static resources from
here(access code:vhm7), and put them on
~/anaconda3/lib/python3.6/site-packages/visdom/static/.

Test Image

python test_image.py

optional arguments:
--upscale_factor      super resolution upscale factor [default value is 3]
--model_name          super resolution model name [default value is epoch_3_100.pt]

The output high resolution images are on results directory.

Test Video

python test_video.py

optional arguments:
--upscale_factor      super resolution upscale factor [default value is 3]
--is_real_time        super resolution real time to show [default value is False]
--delay_time          super resolution delay time to show [default value is 1]
--model_name          super resolution model name [default value is epoch_3_100.pt]

The output high resolution videos are on results directory.

Benchmarks

Adam optimizer were used with learning rate scheduling between epoch 30 and epoch 80.

Upscale Factor = 2

Epochs with batch size of 64 takes ~1 minute on a NVIDIA GeForce TITAN X GPU.

Loss/PSNR graphs

Image Results

The left is low resolution image, the middle is high resolution image, and
the right is super resolution image(output of the ESPCN).

  • Set5
  • Set14
  • BSD100
  • Urban100

Video Results

The right is low resolution video, the left is super resolution video(output of the ESPCN).
Click the image to watch the complete video.

Watch the video

Upscale Factor = 3

Epochs with batch size of 64 takes ~30 seconds on a NVIDIA GeForce TITAN X GPU.

Loss/PSNR graphs

Image Results

The left is low resolution image, the middle is high resolution image, and
the right is super resolution image(output of the ESPCN).

  • Set5
  • Set14
  • BSD100

Video Results

The right is low resolution video, the left is super resolution video(output of the ESPCN).
Click the image to watch the complete video.

Watch the video

Upscale Factor = 4

Epochs with batch size of 64 takes ~20 seconds on a NVIDIA GeForce GTX 1070 GPU.

Loss/PSNR graphs

Image Results

The left is low resolution image, the middle is high resolution image, and
the right is super resolution image(output of the ESPCN).

  • Set5
  • Set14
  • BSD100
  • Urban100

Video Results

The right is low resolution video, the left is super resolution video(output of the ESPCN).
Click the image to watch the complete video.

GitHub