1. Linear Filtering – Part II Selim Aksoy Department of Computer Engineering Bilkent University saksoy@cs.bilkent.edu.tr

2. CS 484, Spring 2010©2010, Selim Aksoy2 Fourier theory Jean Baptiste Joseph Fourier had a crazy idea: Any periodic function can be written as a weighted sum of sines and cosines of different frequencies (1807). Don’t believe it? Neither did Lagrange, Laplace, Poisson, … But it is true!  Fourier series Even functions that are not periodic (but whose area under the curve is finite) can be expressed as the integral of sines and cosines multiplied by a weighing function.  Fourier transform

3. CS 484, Spring 2010©2010, Selim Aksoy3 Fourier theory The Fourier theory shows how most real functions can be represented in terms of a basis of sinusoids. The building block: A sin( ωx + Φ ) Add enough of them to get any signal you want. Adapted from Alexei Efros, CMU

4. CS 484, Spring 2010©2010, Selim Aksoy4 Fourier transform

5. CS 484, Spring 2010©2010, Selim Aksoy5 Fourier transform

6. CS 484, Spring 2010©2010, Selim Aksoy6 Fourier transform

7. CS 484, Spring 2010©2010, Selim Aksoy7 Fourier transform

8. CS 484, Spring 2010©2010, Selim Aksoy8 Fourier transform

9. CS 484, Spring 2010©2010, Selim Aksoy9 Fourier transform

10. CS 484, Spring 2010©2010, Selim Aksoy10 Fourier transform Adapted from Alexei Efros, CMU

11. CS 484, Spring 2010©2010, Selim Aksoy11 Fourier transform Adapted from Gonzales and Woods

12. CS 484, Spring 2010©2010, Selim Aksoy12 Fourier transform Adapted from Gonzales and Woods

13. CS 484, Spring 2010©2010, Selim Aksoy13 Fourier transform

14. CS 484, Spring 2010©2010, Selim Aksoy14 Fourier transform

15. CS 484, Spring 2010©2010, Selim Aksoy15 Fourier transform Adapted from Shapiro and Stockman

16. CS 484, Spring 2010©2010, Selim Aksoy16 Fourier transform Example building patterns in a satellite image and their Fourier spectrum.

17. CS 484, Spring 2010©2010, Selim Aksoy17 Convolution theorem

18. CS 484, Spring 2010©2010, Selim Aksoy18 Frequency domain filtering Adapted from Shapiro and Stockman, and Gonzales and Woods

19. CS 484, Spring 2010©2010, Selim Aksoy19 Frequency domain filtering Since the discrete Fourier transform is periodic, padding is needed in the implementation to avoid aliasing (see section 4.6 in the Gonzales-Woods book for implementation details).

20. CS 484, Spring 2010©2010, Selim Aksoy20 Frequency domain filtering f(x,y) h(x,y) g(x,y)   |F(u,v)| |H(u,v)| |G(u,v)|   Adapted from Alexei Efros, CMU

21. CS 484, Spring 2010©2010, Selim Aksoy21 Smoothing frequency domain filters Adapted from Gonzales and Woods

22. CS 484, Spring 2010©2010, Selim Aksoy22 Smoothing frequency domain filters

23. CS 484, Spring 2010©2010, Selim Aksoy23 Smoothing frequency domain filters The blurring and ringing caused by the ideal low- pass filter can be explained using the convolution theorem where the spatial representation of a filter is given below.

24. CS 484, Spring 2010©2010, Selim Aksoy24 Smoothing frequency domain filters Adapted from Gonzales and Woods

25. CS 484, Spring 2010©2010, Selim Aksoy25 Smoothing frequency domain filters

26. CS 484, Spring 2010©2010, Selim Aksoy26 Sharpening frequency domain filters

27. CS 484, Spring 2010©2010, Selim Aksoy27 Sharpening frequency domain filters Adapted from Gonzales and Woods

28. CS 484, Spring 2010©2010, Selim Aksoy28 Sharpening frequency domain filters Adapted from Gonzales and Woods

29. CS 484, Spring 2010©2010, Selim Aksoy29 Sharpening frequency domain filters Adapted from Gonzales and Woods

30. CS 484, Spring 2010©2010, Selim Aksoy30 Template matching Correlation can also be used for matching. If we want to determine whether an image f contains a particular object, we let h be that object (also called a template) and compute the correlation between f and h. If there is a match, the correlation will be maximum at the location where h finds a correspondence in f. Preprocessing such as scaling and alignment is necessary in most practical applications.

31. CS 484, Spring 2010©2010, Selim Aksoy31 Template matching Adapted from Gonzales and Woods

32. CS 484, Spring 2010©2010, Selim Aksoy32 Template matching Face detection using template matching: face templates.

33. CS 484, Spring 2010©2010, Selim Aksoy33 Template matching Face detection using template matching: detected faces.

34. CS 484, Spring 2010©2010, Selim Aksoy34 Resizing images How can we generate a half-sized version of a large image? Adapted from Steve Seitz, U of Washington

35. CS 484, Spring 2010©2010, Selim Aksoy35 Resizing images Throw away every other row and column to create a 1/2 size image (also called sub-sampling). 1/4 1/8 Adapted from Steve Seitz, U of Washington

36. CS 484, Spring 2010©2010, Selim Aksoy36 Resizing images Does this look nice? 1/4 (2x zoom)1/8 (4x zoom)1/2 Adapted from Steve Seitz, U of Washington

37. CS 484, Spring 2010©2010, Selim Aksoy37 Resizing images We cannot shrink an image by simply taking every k’th pixel. Solution: smooth the image, then sub-sample. Gaussian 1/4 Gaussian 1/8 Gaussian 1/2 Adapted from Steve Seitz, U of Washington

38. CS 484, Spring 2010©2010, Selim Aksoy38 Resizing images Gaussian 1/4 (2x zoom) Gaussian 1/8 (4x zoom) Gaussian 1/2 Adapted from Steve Seitz, U of Washington

39. CS 484, Spring 2010©2010, Selim Aksoy39 Sampling and aliasing Adapted from Steve Seitz, U of Washington

40. CS 484, Spring 2010©2010, Selim Aksoy40 Sampling and aliasing Errors appear if we do not sample properly. Common phenomenon: High spatial frequency components of the image appear as low spatial frequency components. Examples: Wagon wheels rolling the wrong way in movies. Checkerboards misrepresented in ray tracing. Striped shirts look funny on color television.

41. CS 484, Spring 2010©2010, Selim Aksoy41 Gaussian pyramids Adapted from Gonzales and Woods

42. CS 484, Spring 2010©2010, Selim Aksoy42 Gaussian pyramids Adapted from Michael Black, Brown University

43. CS 484, Spring 2010©2010, Selim Aksoy43 Gaussian pyramids Adapted from Michael Black, Brown University