• Stars
    star
    596
  • Rank 75,095 (Top 2 %)
  • Language
    MATLAB
  • Created over 9 years ago
  • Updated over 8 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Single Image Super-Resolution from Transformed Self-Exemplars (CVPR 2015)

Single Image Super-Resolution from Transformed Self-Exemplars (CVPR 2015)

Introduction

This is the research code for the paper:

Jia-Bin Huang, Abhishek Singh, and [Narendra Ahuja] (http://vision.ai.illinois.edu/ahuja.html), "Single Image Super-Resolution from Transformed Self-Exemplars", CVPR 2015 PDF

The proposed algorithm achieves the state-of-the-art performance on image super-resolution without the need of any external training dataset, feature extraction and complicated learning algorithms. For more details, please visit our Project page.

All the datasets (Set5, Set14, Urban 100, BSD 100, Sun-Hays 80), precomputed results and visual comparisons can be found in the following sections.

Citation

If you find the code and dataset useful in your research, please consider citing:

@inproceedings{Huang-CVPR-2015,
    title={Single Image Super-Resolution From Transformed Self-Exemplars},
    Author = {Huang, Jia-Bin and Singh, Abhishek and Ahuja, Narendra},
    booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    pages={5197--5206},
    Year = {2015}
}

Contents

Folder description
cache cached data for vanishing point detection
data Testing images of five datasets (Set5, Set14, Urban 100, BSD 100, Sun-Hays 80). All the images have been cropped according to the desired super-resolution factor. This avoids misalignment of the groundtruth high-resolution images and the super-resolved images
external We use the vgg_interp2 from imrender to perform bilinear interpolation
quant_eval Quantitative evaluation code
reference A copy of the CVPR paper and the bibtex
source MATLAB source code

To run the algorithm on all datasets, simply run the sr_demo_bacth.m. Note that it is an educational code that is not optimized for speed. If timing is a concern, you can achieve visually similar results with small numbers of iterations, e.g., set the number of iterations opt.numIter = 5; in the file sr_init_opt.m. An example of the speed and quality trade-off can be found in Fig. 10 in the paper.

Feedbacks and comments are welcome! Feel free to contact me via [email protected].

Enjoy!

Note: For vanishing point detection only windows executable is provided (from Image Completion using Planar Structure Guidance), a cross-platform version will be included later.

Comparison with the state-of-the-art

Datasets

The full super-resolution results on Set 5, Set 14, Urban 100, BSD 100 and Sun-Hays 80 are available.

| Dataset | Image source | Download full results | |---- | ---|----| ----| | Set 5 | Bevilacqua et al. BMVC 2012 | link (16.1 MB)| | Set 14 | Zeyde et al. LNCS 2010 | link (86.0 MB)| | Urban 100 | Huang et al. CVPR 2015 | link (1.14 GB)| | BSD 100 | Martin et al. ICCV 2001 | link (568 MB)| | Sun-Hays 80 | Sun and Hays ICCP 2012 | link (311 MB)|

Set 5 dataset - link Set 5

Set 14 dataset - link Set 14

Urban 100 dataset - link Urban 100

BSD 100 dataset - link BSD 100

Sun-Hays 80 dataset - link Sun-Hays 80

State-of-the-art image super-resolution algorithms

In each dataset, we include results of the state-of-the-art single image super-resolution algorithms:

Image Description
HR High-resolution images. All images were cropped so that each dimension is a multiplication of the super-resolution factor. This avoids the misalignment problem in the quantitative comparison.
LR Low-resolution test images generated with bicubic kernel downsampling.
bicubic Bicubic interpolation
nearest Nearest-neighbor interpolation
SelfExSR Our result
A+ R. Timofte, V. De Smet, and L. Van Gool, A+: Adjusted Anchored Neighborhood Regression for Fast Super-Resolution. In Asian Conference on Computer Vision (ACCV 2014). Code available here
Abhishek A. Singh and N. Ahuja, Super-Resolution Using Sub-Band Self-Similarity. In Asian Conference on Computer Vision (ACCV 2014). No publicly implementation available. Results were provided by the authors.
Kim K. I. Kim and Y. Kwon, โ€œSingle-image super-resolution using sparse regression and natural image priorโ€, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 32, no. 6, pp. 1127-1133, 2010. Code available here
Glasner Daniel Glasner, Shai Bagon, Michal Irani. Super-Resolution From a Single Image, In International Conference on Computer Vision (ICCV 2009). No public implementation available. Results were generated by our own implementation.
ScSR Jianchao Yang, John Wright, Thomas Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE Transactions on Image Processing, Vol 19, Issue 11, pp2861-2873, 2010. Code available here
SRCNN Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang. Learning a Deep Convolutional Network for Image Super-Resolution, in European Conference on Computer Vision (ECCV 2014). Code available here

Qualitative comparison

In our supplementary material, we includde 120 sample comparisons with the state-of-the-art algorithms. Download the document here.

You can browse and compare our results with other methods via the following links.

Quantitative comparisons

We report three types of metrics

Results on Set 5
Scale Bicubic ScSR Kim Sub-band Glasner SRCNN A+ Ours
2x - PSNR 33.64 35.78 36.24 Sub-band 35.43 36.28 A+ 36.50
3x - PSNR 30.39 31.34 32.30 Sub-band 31.10 32.37 A+ 32.62
4x - PSNR 28.42 29.07 30.07 Sub-band 28.84 30.08 A+ 30.33
2x - SSIM 0.9292 0.9485 0.9518 Sub-band 0.9452 0.9509 A+ 0.9537
3x - SSIM 0.8678 0.8869 0.9041 Sub-band 0.8811 0.9025 A+ 0.9094
4x - SSIM 0.8101 0.8263 0.8553 Sub-band 0.8210 0.8525 A+ 0.8623
2x - IFC 5.72 6.94 7.05 Sub-band 6.70 6.85 A+ 7.83
3x - IFC 3.45 3.98 4.25 Sub-band 3.68 4.11 A+ 4.76
4x - IFC 2.28 2.57 2.82 Sub-band 2.42 2.76 A+ 3.19
Results on Set 14
Scale Bicubic ScSR Kim Sub-band Glasner SRCNN A+ Ours
2x - PSNR 30.22 31.64 32.14 Sub-band 31.41 32.00 A+ 32.23
3x - PSNR 27.53 28.19 28.96 Sub-band 28.21 28.90 A+ 29.16
4x - PSNR 25.99 26.40 27.18 Sub-band 26.43 27.13 A+ 27.40
2x - SSIM 0.8683 0.8940 0.9031 Sub-band 0.8881 0.9012 A+ 0.9036
3x - SSIM 0.7737 0.7977 0.8140 Sub-band 0.7926 0.8124 A+ 0.8197
4x - SSIM 0.7023 0.7218 0.7434 Sub-band 0.7163 0.7395 A+ 0.7518
2x - IFC 5.74 6.83 6.92 Sub-band 6.47 6.68 A+ 7.60
3x - IFC 3.33 3.75 3.92 Sub-band 3.59 3.81 A+ 4.38
4x - IFC 2.18 2.46 2.57 Sub-band 2.30 2.50 A+ 2.90
Results on Urban 100
Scale Bicubic ScSR Kim Sub-band Glasner SRCNN A+ Ours
2x - PSNR 26.66 28.26 28.74 28.34 27.85 28.65 28.87 29.38
4x - PSNR 23.14 24.02 24.20 24.19 23.58 24.14 24.34 24.82
2x - SSIM 0.8408 0.8828 0.8940 0.8820 0.8709 0.8909 0.8957 0.9032
4x - SSIM 0.6573 0.7024 0.7104 0.7115 0.6736 0.7047 0.7195 0.7386
2x - IFC 5.72 6.98 6.86 7.08 6.17 6.66 8.02 7.96
4x - IFC 2.27 2.75 2.71 2.72 2.35 2.63 3.16 3.33
Results on BSD 100
Scale Bicubic ScSR Kim Sub-band Glasner SRCNN A+ Ours
2x - PSNR 29.55 30.77 31.11 30.73 30.28 31.11 31.22 31.18
3x - PSNR 27.20 27.72 28.17 27.88 27.06 28.20 28.30 28.30
4x - PSNR 25.96 26.61 26.71 26.60 26.17 26.70 26.82 26.85
2x - SSIM 0.8425 0.8744 0.8840 0.8774 0.8621 0.8835 0.8862 0.8855
3x - SSIM 0.7382 0.7647 0.7788 0.7714 0.7368 0.7794 0.7836 0.7843
4x - SSIM 0.6672 0.6983 0.7027 0.7021 0.6747 0.7018 0.7089 0.7108
2x - IFC 5.26 6.20 6.30 6.36 5.56 6.09 7.15 6.84
3x - IFC 3.00 3.37 3.49 3.17 2.72 3.39 3.92 3.81
4x - IFC 1.91 2.22 2.20 2.18 1.86 2.18 2.51 2.46