EPSR (Enhanced Perceptual Super-resolution Network) paper

This repo provides the test code, pretrained models, and results on benchmark datasets of our work. We (IPCV_team) won the first place in PIRM2018-SR competition (region 1). We were also ranked as second and thrid in region 2 and 3 respectively. For details refer to our recently accepted paper in ECCV2018 PIRM Workshop.

"Analyzing Perception-Distortion Tradeoff using Enhanced Perceptual Super-resolution Network", Subeesh Vasu, Nimisha T. M. and A. N. Rajagopalan, Perceptual Image Restoration and Manipulation (PIRM) Workshop and Challenge, Eurpean Conference on Computer Vision Workshops (ECCVW 2018), Munich, Germany, September 2018. [arXiv]

BibTeX

 @inproceedings{vasu2018analyzing,
    author = {Vasu, Subeesh and T.M., Nimisha and Rajagopalan, A.N.},
    title = {Analyzing Perception-Distortion Tradeoff using Enhanced Perceptual Super-resolution Network},
    booktitle = {European Conference on Computer Vision (ECCV) Workshops},
    year = {2018}}

Results

Visual comparison for 4× SR with bicubic interpolation model on PIRM-self, BSD100, and Urban100 datasets. Here I_HR refers to the ground truth HR image. SRCNN, EDSR, DBPN, ENet, and CX are existing works. EPSR₁, EPSR₂, and EPSR₃ are the results of our approach (EPSR) corresponding to region 1, 2, and 3 of PIRM-SR challenge. BNet₁, BNet₂, and BNet₃ are the results of our baseline network.

Perception-distortion trade-off between BNet and EPSR. For both methods, the above plot has the values corresponding to 19 model weights which span different regions on the perception-distortion plane and the corresponding curves that best fit these values.

Performance comparison of top 9 methods from PIRM-SR challenge. Methods are ranked based on the PI and RMSE values corresponding to the test data of PIRM-SR. The entries from our approach are highlighted in red. Methods with a marginal difference in PI and RMSE values share the same rank and are indicated with a " * ".

Test

The code is built on the official implementation of EDSR (PyTorch) and tested on Ubuntu 16.04 environment (Python3.6, PyTorch_0.4.0, CUDA8.0) with Titan X GPU. Refer EDSR (PyTorch) for other dependencies. Test code of EPSR can be found in EPSR_testcode.

Results on public benchmark datasets

• PIRM-self
• PIRM-test
• BSD100
• Urban100
• Set14
• Set5

References

[SRCNN] Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. ECCV 2014

[EDSR] Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. CVPR workshops 2017

[DBPN] Haris, M., Shakhnarovich, G., Ukita, N.: Deep backprojection networks for super-resolution. CVPR 2018

[ENet] Sajjadi, M.S., Sch ̈olkopf, B., Hirsch, M.: Enhancenet: Single image super-resolution through automated texture synthesis. ICCV 2017

[CX] Mechrez, R., Talmi, I., Shama, F., Zelnik-Manor, L. Learning to maintain natural image statistics. arXiv preprint arXiv:1803.04626 (2018)

[PIRM-SR challenge] Blau, Y., Mechrez, R., Timofte, R. 2018 PIRM Challenge on Perceptual Image Super-resolution. arXiv preprint arXiv:1809.07517 (2018)

Acknowledgements

This code is built on EDSR (PyTorch). We thank the authors for sharing their codes of EDSR PyTorch version.