Super-Resolution.Benckmark

A curated list of super-resolution resources and a benchmark for single image super-resolution algorithms.

See my implementated super-resolution algorithms:

• SRGAN
• VDSR
• CSCN

TODO

Build a benckmark like SelfExSR_Code

State-of-the-art algorithms

Classical Sparse Coding Method

• ScSR [Web]
• Image super-resolution as sparse representation of raw image patches (CVPR2008), Jianchao Yang et al.
• Image super-resolution via sparse representation (TIP2010), Jianchao Yang et al.
• Coupled dictionary training for image super-resolution (TIP2011), Jianchao Yang et al.

Anchored Neighborhood Regression Method

• ANR [Web]
• Anchored Neighborhood Regression for Fast Example-Based Super-Resolution (ICCV2013), Radu Timofte et al.
• A+ [Web]
• A+: Adjusted Anchored Neighborhood Regression for Fast Super-Resolution (ACCV2014), Radu Timofte et al.
• IA [Web]
• Seven ways to improve example-based single image super resolution (CVPR2016), Radu Timofte et al.

Self-Exemplars

• SelfExSR [Web]
• Single Image Super-Resolution from Transformed Self-Exemplars (CVPR2015), Jia-Bin Huang et al.

Bayes

• NBSRF [Web]
• Naive Bayes Super-Resolution Forest (ICCV2015), Jordi Salvador et al.

Deep Learning Method

• SRCNN [Web] [waifu2x by nagadomi]
• Image Super-Resolution Using Deep Convolutional Networks (ECCV2014), Chao Dong et al.
• Image Super-Resolution Using Deep Convolutional Networks (TPAMI2015), Chao Dong et al.
• CSCN [Web]
• Deep Networks for Image Super-Resolution with Sparse Prior (ICCV2015), Zhaowen Wang et al.
• Robust Single Image Super-Resolution via Deep Networks with Sparse Prior (TIP2016), Ding Liu et al.
• VDSR [Web] [Unofficial Implementation in Caffe]
• Accurate Image Super-Resolution Using Very Deep Convolutional Networks (CVPR2016), Jiwon Kim et al.
• DRCN [Web]
• Deeply-Recursive Convolutional Network for Image Super-Resolution (CVPR2016), Jiwon Kim et al.
• ESPCN [PDF]
• Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network (CVPR2016), Wenzhe Shi et al.
• Is the deconvolution layer the same as a convolutional layer? [PDF]
• FSRCNN [Web]
• Acclerating the Super-Resolution Convolutional Neural Network (ECCV2016), Dong Chao et al.

Perceptual Loss and GAN

• Perceptual Loss [PDF]
• Perceptual Losses for Real-Time Style Transfer and Super-Resolution (ECCV2016), Justin Johnson et al.
• SRGAN [PDF]
• Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, Christian Ledig et al.
• AffGAN [PDF]
• AMORTISED MAP INFERENCE FOR IMAGE SUPER-RESOLUTION, Casper Kaae Sønderby et al.
• EnhanceNet [PDF]
• EnhanceNet: Single Image Super-Resolution through Automated Texture Synthesis, Mehdi S. M. Sajjadi et al.
• neural-enchance [Github]

Video SR

• VESPCN [[PDF]](Real-Time Video Super-Resolution with Spatio-Temporal Networks and Motion Compensation)

Dicussion

Deconvolution and Sub-Pixel Convolution

• Deconvolution and Checkerboard Artifacts
• SubPixel

Datasets

| Test Dataset | Image source | |---- | ---|----| | Set 5 | Bevilacqua et al. BMVC 2012 | | Set 14 | Zeyde et al. LNCS 2010 | | BSD 100 | Martin et al. ICCV 2001 | | Urban 100 | Huang et al. CVPR 2015 |

| Train Dataset | Image source | |---- | ---|----| | Yang 91 | Yang et al. CVPR 2008 | | BSD 200 | Martin et al. ICCV 2001 | | General 100 | Dong et al. ECCV 2016 | | ImageNet | Olga Russakovsky et al. IJCV 2015 | | COCO| Tsung-Yi Lin et al. ECCV 2014

Quantitative comparisons

Results from papers of VDSR, DRCN, CSCN and IA.

Note: IA use enchanced prediction trick to improve result.

Results on Set 5

Scale	Bicubic	A+	SRCNN	SelfExSR	CSCN	VDSR	DRCN	IA
2x - PSNR/SSIM	33.66/0.9929	36.54/0.9544	36.66/0.9542	36.49/0.9537	36.93/0.9552	37.53/0.9587	37.63/0.9588	37.39/
3x - PSNR/SSIM	30.39/0.8682	32.59/0.9088	32.75/0.9090	32.58/0.9093	33.10/0.9144	33.66/0.9213	33.82/0.9226	33.46/
4x - PSNR/SSIM	28.42/0.8104	30.28/0.8603	30.48/0.8628	30.31/0.8619	30.86/0.8732	31.35/0.8838	31.53/0.8854	31.10/

Results on Set 14

Scale	Bicubic	A+	SRCNN	SelfExSR	CSCN	VDSR	DRCN	IA
2x - PSNR/SSIM	30.24/0.8688	32.28/0.9056	32.42/0.9063	32.22/0.9034	32.56/0.9074	33.03/0.9124	33.04/0.9118	32.87/
3x - PSNR/SSIM	27.55/0.7742	29.13/0.8188	29.28/0.8209	29.16/0.8196	29.41/0.8238	29.77/0.8314	29.76/0.8311	29.69/
4x - PSNR/SSIM	26.00/0.7027	27.32/0.7491	27.49/0.7503	27.40/0.7518	27.64/0.7587	28.01/0.7674	28.02/0.7670	27.88/

Results on BSD 100

Scale	Bicubic	A+	SRCNN	SelfExSR	CSCN	VDSR	DRCN	IA
2x - PSNR/SSIM	29.56/0.8431	31.21/0.8863	31.36/0.8879	31.18/0.8855	31.40/0.8884	31.90/0.8960	31.85/0.8942	31.79/
3x - PSNR/SSIM	27.21/0.7385	28.29/0.7835	28.41/0.7863	28.29/0.7840	28.50/0.7885	28.82/0.7976	28.80/0.7963	28.76/
4x - PSNR/SSIM	25.96/0.6675	26.82/0.7087	26.90/0.7101	26.84/0.7106	27.03/0.7161	27.29/0.7251	27.23/0.7233	27.25/

-- Results on Set 5 (PSNR/SSIM/Time)

| Scale | Bicubic | A+ | SRCNN | SelfExSR | SCN | VDSR | DRCN | PSyCo (32)|PSyCo (1024) | FSRCNN-S | FSRCNN | RAISR |
|:---------:|:-------:|:--------:|:------:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:| | 2x | 33.66/0.002 | 36.54/0.684 | 36.66/4.722 | 36.49/42.521 | 36.93 | 37.53/0.9587/0.13 | 37.63/0.9588/1.54 | 36.57/0.038 | 36.88/0.185 | 36.58/0.024 | 37.00/0.068 |36.061/0.951/0.018 |
| 3x | 30.39/0.002 | 32.59/0.401 | 32.75/5.226 | 32.58/31.008 | 33.10 | 33.66/0.9213/0.13 | 33.82/0.9226/1.55 | 32.63/0.049 | 32.93/0.456 | 32.61/0.010 | 33.16/0.027 |32.172/0.900/0.015 | | 4x | 28.42/0.002 | 30.28/0.226 | 30.48/9.962 | 30.31/26.728 | 30.86 | 31.35/0.8838/0.12 | 31.53/0.8854/1.54 | 30.32/0.055 | 30.62/0.210 | 30.11/0.0052 | 30.71/0.015 |29.834/0.848/0.017 |

---

[ RAISR ] RAISR: Rapid and Accurate Image Super Resolution

[ PSyCo ] PSyCo: Manifold Span Reduction for Super Resolution

[ FSRCNN ] Accelerating the Super-Resolution Convolutional Neural Network

[ DRCN ] Deeply-Recursive Convolutional Network for Image Super-Resolution

[ AI ] Seven ways to improve example-based single image super resolution

[ VDSR ] Accurate Image Super-Resolution Using Very Deep Convolutional Networks

[ SCN ] Deep Networks for Image Super-Resolution with Sparse Prior

[ SRCNN ] Image Super-Resolution Using Deep Convolutional Networks

[ ANR ] Anchored Neighborhood Regression for Fast Example-Based Super-Resolution

[ A+ ] A+: Adjusted Anchored Neighborhood Regression for Fast Super-Resolution