Sharpening Out of Focus Images using High-Frequency Transfer
Paper ID: 1012
Comparisons Supplementary

We compare our algorithm against sampled previous works in deconvolution, super resolution, image processing, and motion-deblurring. Works include Lucy-Richardson, Levin et al. 2007, Krishnan and Fergus 2009, Glasner et al. 2009, Yang et al. 2010, Freedman and Fattal 2011, and unsharp mask.

Instructions:
Click on the thumbnails to view the full image in a new tab.

Suggestions:
- The brighter the screen, the better

Notes:
- Some translation shifts are completely normal- algorithms process with different padding parameters, etc.
- Glasner et al. 2009 and Freedman and Fattal 2011 were run on two of the examples because their codes are not publically available. From our observations, the algorithms perform similarly to Yang et al. 2010. We would like to thank Daniel Glasner and Gilal Freedman for running their codes on the two examples for us.

Real professional sports photographer's example

Notes: 5D Mark III Look for: - Skin detail, face sharpness, text sharpness on shirt Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Real portrait photographer's example

Notes: 1D Mark IV, ISO 250 Look for: - Details of wrinkles, naturalness of the wrinkles Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Real wildlife photographer's example

Notes: 5D Mark II, ISO 100, missed focus Look for: - Our method reduces circle of confusion while others emphasize the edges - Fur on the deer - Noise levels (Unsharp mask emphasizes noise) Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Real bird photographer's example

Notes: - Small amount of motion blur Look for: - Feather details Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Glasner et al. 2009) Super-Resolution (Yang et al. 2010) Super-Resolution (Freedman and Fattal 2011) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Glasner et al. 2009 Yang et al. 2010 Freedman and Fattal 2011 Unsharp Mask

Real professional sports photographer's example

Notes: 5D Mark III Look for: - Text detail - Noise levels (Unsharp mask emphasizes noise) Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Real bird photographer's example

Look for: - Eyes - Dark feather details Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Glasner et al. 2009) Super-Resolution (Yang et al. 2010) Super-Resolution (Freedman and Fattal 2011) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Glasner et al. 2009 Yang et al. 2010 Freedman and Fattal 2011 Unsharp Mask

Real high ISO bird photography example

Notes: 5D Mark II, ISO 3200 Look for: - Details of fur and head Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Real large blur kernel example (failure case for all methods)

Notes: 5D Mark II, large aperture at f/1.4, completely missed focus (large blur kernel) - Because of the large blur kernel, although all algorithms perform non-ideally- our algorithm produces the best result Look for: - Details of glasses Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Real professional sports photographer's example (Higher Resolution 1200 x 1200)

Notes: 5D Mark III Look for: - Details of cap, face Assisting Image Out of Focus Image Our Result Out of Focus Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Unsharp Mask Full-Sized Comparison (click links on the right to view image)                 (Click here to view image) Out of focus Input Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Unsharp Mask

Synthetically blurred examples

Look for: - Details of grass and helmet Assisting Image Out of Focus Image Ground Truth Our Result Out of Focus Ground Truth Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Sparsity Prior (Levin et al. 2007) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Ground Truth Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Levin et al. 2007 Unsharp Mask

Look for: - Overall sharpness and license plate Assisting Image Out of Focus Image Ground Truth Our Result Out of Focus Ground Truth Our Result Lucy-Richardson Motion-Deblur (Krishnan and Fergus 2009) Super-Resolution (Yang et al. 2010) Sparsity Prior (Levin et al. 2007) Unsharp Mask Full-Sized Comparison (click links on the right to view image) (Click here to view image) Out of focus Input Ground Truth Our Result Lucy-Richardson Krishnan and Fergus 2009 Yang et al. 2010 Levin et al. 2007 Unsharp Mask

Depth-of-Field Expansion

In this scenario, the user focuses in the background and foreground with different exposures for the assisting images (a). For the target image (b), both background and foreground are out-of-focus. With our algorithm, we output a result that is sharper in both the foreground and background (c). Notice that the exposure and color is retained in the output.

References

FREEDMAN, G., AND FATTAL, R. 2011. Image and video upscaling from local self-examples. ACM TOG.

GLASNER, D., BAGON, S., AND IRANI, M. 2009. Super-resolution from a single image. ICCV.

KRISHNAN, D., AND FERGUS, R. 2009. Fast image deconvolution using hyper-laplacian priors. In Proc. of Neural Information Processing Systems.

LEVIN, A., FERGUS, R., DURAND, F., AND FREEMEN, W. T. 2007. Image and depth from a conventional camera with a coded aperture. ACM SIGGRAPH.

YANG, J., WRIGHT, J., HUANG, T., AND MA, Y. 2010. Image super resolution via sparse representation. IEEE Transactions on Image Processing.

Parameter Notes

Our Method
- We use the same parameters to generate all the examples

Freedman and Fattal 2011
- Factors of 5/4*5/4*4/3*3/2 and downsampled the result from factor of 3.125 to factor of 3
- L1 distance function in the patch search

Unsharp Mask (Photoshop)
- 100% sharpen, 2 pixels, 0 level tolerance

All other methods use default parameters given by the authors.