This repository contains an op-for-op PyTorch reimplementation of Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network.
- ESPCN-PyTorch
About Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network
If you're new to ESPCN, here's an abstract straight from the paper:
Recently, several models based on deep neural networks have achieved great success in terms of both reconstruction accuracy and computational performance for single image super-resolution. In these methods, the low resolution (LR) input image is upscaled to the high resolution (HR) space using a single filter, commonly bicubic interpolation, before reconstruction. This means that the super-resolution (SR) operation is performed in HR space. We demonstrate that this is sub-optimal and adds computational complexity. In this paper, we present the first convolutional neural network (CNN) capable of real-time SR of 1080p videos on a single K2 GPU. To achieve this, we propose a novel CNN architecture where the feature maps are extracted in the LR space. In addition, we introduce an efficient sub-pixel convolution layer which learns an array of upscaling filters to upscale the final LR feature maps into the HR output. By doing so, we effectively replace the handcrafted bicubic filter in the SR pipeline with more complex upscaling filters specifically trained for each feature map, whilst also reducing the computational complexity of the overall SR operation. We evaluate the proposed approach using images and videos from publicly available datasets and show that it performs significantly better (+0.15dB on Images and +0.39dB on Videos) and is an order of magnitude faster than previous CNN-based methods.
Contains DIV2K, DIV8K, Flickr2K, OST, T91, Set5, Set14, BSDS100 and BSDS200, etc.
Please refer to README.md in the data directory for the method of making a dataset.
Both training and testing only need to modify the config.py file.
Modify the config.py file.
- line 31:
model_arch_namechange toespcn_x4. - line 36:
upscale_factorchange to4. - line 38:
modechange totest. - line 40:
exp_namechange toESPCN_x4-Set5. - line 84:
lr_dirchange tof"./data/Set5/LRbicx{upscale_factor}". - line 86:
gt_dirchange tof"./data/Set5/GTmod12". - line 88:
model_weights_pathchange to./results/pretrained_models/ESPCN_x4-T91-64bf5ee4.pth.tar.
python3 test.pyModify the config.py file.
- line 31:
model_arch_namechange toespcn_x4. - line 36:
upscale_factorchange to4. - line 38:
modechange totest. - line 40:
exp_namechange toESPCN_x4-Set5. - line 84:
lr_dirchange tof"./data/Set5/LRbicx{upscale_factor}". - line 86:
gt_dirchange tof"./data/Set5/GTmod12".
python3 train.pyModify the config.py file.
- line 31:
model_arch_namechange toespcn_x4. - line 36:
upscale_factorchange to4. - line 38:
modechange totest. - line 40:
exp_namechange toESPCN_x4-Set5. - line 57:
resume_model_weights_pathchange to./samples/ESPCN_x4-Set5/epoch_xxx.pth.tar. - line 84:
lr_dirchange tof"./data/Set5/LRbicx{upscale_factor}". - line 86:
gt_dirchange tof"./data/Set5/GTmod12".
python3 train.pySource of original paper results: https://arxiv.org/pdf/1609.05158v2.pdf
In the following table, the value in () indicates the result of the project, and - indicates no test.
| Method | Scale | Set5 (PSNR) | Set14 (PSNR) |
|---|---|---|---|
| ESPCN_x4 | 2 | -(36.64) | -(32.35) |
| ESPCN_x3 | 3 | 32.55(32.55) | 29.08(29.20) |
| ESPCN_x4 | 4 | 30.90(30.26) | 27.73(27.41) |
# Download `ESPCN_x4-T91-64bf5ee4.pth.tar` weights to `./results/pretrained_models/ESPCN_x4-T91-64bf5ee4.pth.tar`
# More detail see `README.md<Download weights>`
python3 ./inference.pyInput:
Output:
Build `espcn_x4` model successfully.
Load `espcn_x4` model weights `./results/pretrained_models/ESPCN_x4-T91-64bf5ee4.pth.tar` successfully.
SR image save to `./figure/sr_comic.png`
Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network
Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, Zehan Wang
Abstract
Recently, several models based on deep neural networks have achieved great success in terms of both reconstruction accuracy and computational
performance for single image super-resolution. In these methods, the low resolution (LR)
input image is upscaled to the high resolution (HR) space using a single filter, commonly bicubic interpolation, before reconstruction. This means
that the super-resolution (SR) operation is performed in HR space. We demonstrate that this is sub-optimal and adds computational complexity. In this
paper, we present the first convolutional neural network (CNN) capable of real-time SR of 1080p videos on a single K2 GPU. To achieve this, we propose
a novel CNN architecture where the feature maps are extracted in the LR space. In addition, we introduce an efficient sub-pixel convolution layer
which learns an array of upscaling filters to upscale the final LR feature maps into the HR output. By doing so, we effectively replace the
handcrafted bicubic filter in the SR pipeline with more complex upscaling filters specifically trained for each feature map, whilst also reducing the
computational complexity of the overall SR operation. We evaluate the proposed approach using images and videos from publicly available datasets and
show that it performs significantly better (+0.15dB on Images and +0.39dB on Videos) and is an order of magnitude faster than previous CNN-based
methods.
@article{DBLP:journals/corr/ShiCHTABRW16,
author = {Wenzhe Shi and
Jose Caballero and
Ferenc Husz{\'{a}}r and
Johannes Totz and
Andrew P. Aitken and
Rob Bishop and
Daniel Rueckert and
Zehan Wang},
title = {Real-Time Single Image and Video Super-Resolution Using an Efficient
Sub-Pixel Convolutional Neural Network},
journal = {CoRR},
volume = {abs/1609.05158},
year = {2016},
url = {http://arxiv.org/abs/1609.05158},
archivePrefix = {arXiv},
eprint = {1609.05158},
timestamp = {Mon, 13 Aug 2018 16:47:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/ShiCHTABRW16.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}