This repository contains the code (in PyTorch) for "Efficient Single Image Super Resolution using Enhanced Learned Group Convolutions", paper by Vandit Jain, Prakhar Bansal, Abhinav Kumar Singh, Rajeev Srivastava.
If you find our project useful in your research, please consider citing:
@article{DBLP:journals/corr/abs-1808-08509,
author = {Vandit Jain and
Prakhar Bansal and
Abhinav Kumar Singh and
Rajeev Srivastava},
title = {Efficient Single Image Super Resolution using Enhanced Learned Group
Convolutions},
journal = {CoRR},
volume = {abs/1808.08509},
year = {2018},
url = {http://arxiv.org/abs/1808.08509},
archivePrefix = {arXiv},
eprint = {1808.08509},
timestamp = {Sun, 02 Sep 2018 15:01:55 +0200},
biburl = {https://dblp.org/rec/bib/journals/corr/abs-1808-08509},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
we propose a novel SISR method that uses relatively less number of computations. On training, we get group convolutions that have unused connections removed. We have refined this system specifically for the task at hand by removing unnecessary modules from original CondenseNet. Further, a reconstruction network consisting of deconvolutional layers has been used in order to upscale to high resolution. All these steps significantly reduce the number of computations required at testing time. We evaluate the method using various benchmark datasets and show that it performs favourably against the state-of-the-art methods in terms of both accuracy and number of computations required.
python main.py --model condensenet -b 256 -j 20 /PATH/TO/DATA \
--stages 7-7-7-7 --growth 14-14-14-14
We have used 91 images from Yang et al. and 200 images from the Berkeley Segmentation Dataset(BSD) with 32x32 patches for training.
Dataset | PSNR | SSIM |
---|---|---|
Set5 | 37.79 | 0.9594 |
Set14 | 33.23 | 0.9137 |
Urban100 | 31.24 | 0.9190 |
Flops count to produce a 64x64 output image comes out to be 668.88 million.