1 / 47

Robust Super-Resolution

Robust Super-Resolution. Presented By: Sina Farsiu. Project Goals. Understanding & simulation of Dr. Assaf Zomet, et.al. paper : “Robust Super-Resolution” [1] Comparing the results obtained by this method to other methods Enhancing the method introduced in: [1] .

claudia
Download Presentation

Robust Super-Resolution

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Robust Super-Resolution Presented By: Sina Farsiu

  2. Project Goals • Understanding & simulation of Dr. Assaf Zomet, et.al. paper : “Robust Super-Resolution” [1] • Comparing the results obtained by this method to other methods • Enhancing the method introduced in: [1]

  3. Super-Resolution Objective • Generation of a high-resolution image from multiple low-resolution moving frames of a scene.

  4. Super-Resolution Formulation

  5. Solutions for S-R Problem • No Noise: • With Noise

  6. Problem with These Solutions • In the presence of outliers (error in motion estimation, inaccurate blur model, pepper & salt noise, …), these methods do not work accurately. • Robust S-R methods can help in these situations.

  7. Robust S-R Formulation

  8. Robust S-R Formulation • Robustness

  9. Why Median? • Median is an estimate of mean. • Unlike regularization method only one of low resolution frames contributes to reconstruct each pixel in the high-resolution frame. So outliers in other frames are discarded in the reconstruction process. • Claim: In the absence of additive noise to all frames median method works better than regularization method.

  10. What if noise is added to all frames? • Claim: If considerable amount of additive noise is present in all frames regularization method works as good or even better than median method.

  11. Median-Average Reconstruction • Instead of using We can combine average and median operators to get better results.

  12. Bias Detection Procedure • It is useful to detect the outliers in the low resolution frames. • We can omit those outlier pixels in our procedures. • The difficulty is to differentiate between aliasing and outlier effects.

  13. Bias Detection • Formulation • After thresholding non zero values are due to aliasing or outliers.

  14. Due to outlier Due to aliasing Bias Detection Result

  15. Bias detection Procedure • 1: compute • 2: Threshold • 3: Filter the result with a LPF • 4: Threshold • 5: Omit corresponding pixels from super-resolution procedure

  16. Bias Detection • B-D method works only for uniform gray level outliers. • In many situations median operation in robust super-resolution eliminates the bias of the estimator.

  17. Original L-R Frame

  18. Robust S-R Reconstructed H-R Frame mse=0.0017

  19. Median Reconstruction after adding noise 0.0131

  20. Regularized S-R Reconstructed H-R Frame mse=0.0131`

  21. Regularized Reconstruction after adding noise mse=.0125

  22. Error Due to Outlier

  23. Error Due to Aliasing

  24. Conclusion • Robust super-resolution method is quite effective in the presence of outliers, and produces better results in comparison with regularization method. • In the presence of additive noise in all low –resolution frames this method loses its superiority to the regularization method.

  25. Conclusion & Results • Combination of mean and median operators can help us in this situation. • Proposed bias detection algorithm is an effective method to detect outliers. • If outliers are the only source of error in the L-R frames(no additive noise), more iterations we use smaller mse we will achieve.

  26. Suggestions for Future Research • Combining regularization and robust super-resolution methods. • Using bias detection results in regularization method. • Using robust super-resolution method in frequency domain.

  27. Acknowledgment • Thanks to Dr.Assaf Zomet, Dr.Michael Elad, Dirk Robinson and Dr. Peyman Milanfar for their valuable advices & suggestions.

  28. references • "Robust Super Resolution",  A. Zomet, A. Rav-Acha and S. Peleg Proceedings of the International Conference on Computer Vision and Pattern Recognition (CVPR), Hawaii, December 2001. • "Efficient Super-Resolution and Applications to Mosaics", A. Zomet and S. Peleg, Proceedings of the International Conference on Pattern Recognition (ICPR), Barcelona, September 2000.

  29. “A Computationally effective Image super-resolution Algorithm”, Nguyen, N., P. Milanfar, G.H. Golub, IEEE Transactions on Image Processing, vol. 10, no. 4, pp. 573-583, April 2001 • “A Fast Super-Resolution Reconstruction Algorithm for Pure Transnational Motion and Common Space Invariant Blur”, M. Elad and Y. Hel-Or, the IEEE Trans. on Image Processing, Vol.10, no.8, pp.1187-93, August 2001.

  30. Thank You All

  31. Additional Simulatins

  32. Original H-R Frame

  33. Blured median

  34. Projected L-R frame

  35. L-R Frame

  36. Regularization with high noisemse=.0481

  37. Median with high noisemse=.0693

  38. Composite Median & AverageMSE=0.0592

  39. Original H.R. Frame

  40. L.R. Frame Before Adding Noise

  41. L.R. Frame After Adding Noise

  42. Regularized ReconstructionMSE=.0216

  43. Median ReconstructionMSE=.0077

More Related