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

Radu Timofte, Vincent De Smet, and Luc Van Gool

ACCV 2014 Submission ID 820

This supplementary material provides codes, images, PSNR results, and running-time values for the methods from the paper:

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), 1-5 November 2014, Singapore. [ bib |  pdf |  Supplementary material |  Source codes ]



A+ is shown to improve both in quality and speed over the compared methods.
The settings are the ones from the paper, the same evaluation procedure as in:
- Timofte et al, ICCV, 2013
- Zeyde et al, 'Curves and Surfaces', 2010

Besides the corresponding results to the ones in the Tables 1,2, and 3, we additionaly provide results for:
1) A+ (0.5 mil.) : our A+ with 1024 atoms, but with 0.5 million mined training samples (from the same training images) for learning the regressors at each upscaling factor
both A+ and A+ (0.5mil.) improve greatly over ANR in each setting (!)
2) A+ (16 atoms): our A+ with 16 atoms (dictionary size = 16), no supplementary mined samples
A+ (16 atoms) improves over ANR, but it uses only 16 regressors instead of 1024, thus a x64 algorithmic speedup (!)

Average performance in PeakSNR [dB], the last column uses a 16 atoms dictionary, all the other use 1024 dictionaries as in Table 3 from the paper.
DatasetScaleBicubicSFZeyde et al.GRANRNE+LSNE+NNLSNE+LLEOur A+ (0.5mil)Our A+ Our A+ (only 16 atoms!)
Set5x233.6635.6335.7835.1335.8335.6635.4335.7736.3736.55 (+0.72 over ANR)35.95
Set5x330.3931.2731.9031.4131.9231.7831.6031.8432.3932.59 (+0.67 over ANR)32.07
Set5x428.4228.9429.6929.3429.6929.5529.4729.6130.0430.29 (+0.60 over ANR)29.83
Set14x230.2331.0431.8131.3631.8031.6931.5531.7632.1332.28 (+0.48 over ANR)31.91
Set14x327.5428.2028.6728.3128.6528.5928.4428.6028.9729.13 (+0.48 over ANR)28.71
Set14x426.0026.2526.8826.6026.8526.8126.7226.8127.1027.32 (+0.47 over ANR)26.99
B100x229.3230.3530.4030.2330.4430.3630.2730.4130.6830.78 (+0.34 over ANR)30.53
B100x327.1527.7627.8727.7027.8927.8327.7327.8528.0828.18 (+0.29 over ANR)27.94
B100x425.9226.1926.5126.3726.5126.4526.4126.4726.6526.77 (+0.26 over ANR)26.58

(Note that it is a re-run, there might be some differences in comparison with the paper.)

Please click on the PSNR result values to see the grayscale image output for most of the methods.


Set5 upscaling x2 results (images, PSNR, times)(Table 2 in the paper)
Set5 upscaling x3 results (images, PSNR, times)(Table 2 in the paper)
Set5 upscaling x4 results (images, PSNR, times)(Table 2 in the paper)
Set14 upscaling x2 results (images, PSNR, times)(Table 3 in the paper)
Set14 upscaling x3 results (images, PSNR, times)(Table 1 in the paper)
Set14 upscaling x4 results (images, PSNR, times)(Table 3 in the paper)
B100 upscaling x2 results (images, PSNR, times)(Table 3 in the paper)
B100 upscaling x3 results (images, PSNR, times)(Table 3 in the paper)
B100 upscaling x4 results (images, PSNR, times)(Table 3 in the paper)

Check out seven ways to improve super-resolution and our new IA method (+0.9dB improvements) !

[43] R. Timofte, R. Rothe, and L. Van Gool. Seven ways to improve example-based single image super resolution. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2016), June 2016, US. [ bib |  pdf |  Supplementary material |  L20 dataset ]