The Frequency Domain, without tears. Many slides borrowed from Steve Seitz

Transcription

1 The Frequency Domain, without tears Many slides borrowed from Steve Seitz Somewhere in Cinque Terre, May 2005 CS194: Image Manipulation & Computational Photography Alexei Efros, UC Berkeley, Fall 2016

2 Salvador Dali Gala Contemplating the Mediterranean Sea, which at 30 meters becomes the portrait of Abraham Lincoln, 1976

3

4

5 A nice set of basis Teases away fast vs. slow changes in the image. This change of basis has a special name

6 Jean Baptiste Joseph Fourier ( ) had crazy idea (1807): Any univariate function can be rewritten as a weighted sum of sines and cosines of different frequencies. Don t believe it? Neither did Lagrange, Laplace, Poisson and other big wigs Not translated into English until 1878! But it s (mostly) true! called Fourier Series...the manner in which the author arrives at these equations is not exempt of difficulties and...his analysis to integrate them still leaves something to be desired on the score of generality and even rigour. Laplace Lagrange Legendre

7 A sum of sines Our building block: Asin(ωx +φ) Add enough of them to get any signal f(x) you want! How many degrees of freedom? What does each control? Which one encodes the coarse vs. fine structure of the signal?

8 Fourier Transform We want to understand the frequency ω of our signal. So, let s reparametrize the signal by ω instead of x: f(x) Fourier Transform F(ω) For every ω from 0 to inf, F(ω) holds the amplitude A and phase φ of the corresponding sine How can F hold both? A = ± R ω + I F(ω) F ( ω) = R( ω) + ii( ω) 2 2 ( ) ( ω) We can always go back: φ = Inverse Fourier Transform tan 1 Asin(ωx +φ) I( ω) R( ω) f(x)

9 Time and Frequency example : g(t) = sin(2pf t) + (1/3)sin(2p(3f) t)

10 Time and Frequency example : g(t) = sin(2pf t) + (1/3)sin(2p(3f) t) = +

11 Frequency Spectra example : g(t) = sin(2pf t) + (1/3)sin(2p(3f) t) = +

12 Frequency Spectra Usually, frequency is more interesting than the phase

13 Frequency Spectra = + =

14 Frequency Spectra = + =

15 Frequency Spectra = + =

16 Frequency Spectra = + =

17 Frequency Spectra = + =

18 Frequency Spectra = A k = 1 1 sin(2 π kt ) k

19 Frequency Spectra

20 FT: Just a change of basis M * f(x) = F(ω) * =...

21 IFT: Just a change of basis M -1 * F(ω) = f(x) * =...

22 Finally: Scary Math

23 Finally: Scary Math not really scary: e i ω x is hiding our old friend: phase can be encoded by sin/cos pair = cos( ωx) + i sin( ωx) sin(ωx +φ) P cos( x) + Qsin( x) = Asin( x + φ) Α = ± P 2 + Q 2 φ = tan 1 P Q So it s just our signal f(x) times sine at frequency ω

24 Extension to 2D = Image as a sum of basis images

25 Extension to 2D in Matlab, check out: imagesc(log(abs(fftshift(fft2(im)))));

26 Fourier analysis in images Intensity Image Fourier Image

27 Signals can be composed + = More:

28 Man-made Scene

29 Can change spectrum, then reconstruct Local change in one domain, courses global change in the other

30 Low and High Pass filtering

31 The Convolution Theorem The greatest thing since sliced (banana) bread! The Fourier transform of the convolution of two functions is the product of their Fourier transforms F[ g h] = F[ g]f[ h] The inverse Fourier transform of the product of two Fourier transforms is the convolution of the two inverse Fourier transforms F [ gh] = F [ g] [ h] Convolution in spatial domain is equivalent to multiplication in frequency domain! F

32 2D convolution theorem example f(x,y) F(s x,s y ) * h(x,y) H(s x,s y ) g(x,y) G(s x,s y )

33 Filtering Why does the Gaussian give a nice smooth image, but the square filter give edgy artifacts? Gaussian Box filter

34 Fourier Transform pairs

35 Gaussian

36 Box Filter

37 Low-pass, Band-pass, High-pass filters low-pass: High-pass / band-pass:

38 Edges in images

39 What does blurring take away? original

40 What does blurring take away? smoothed (5x5 Gaussian)

41 High-Pass filter smoothed original

42 Image Sharpening What does blurring take away? = original smoothed (5x5) detail Let s add it back: + α = original detail sharpened

43 Unsharp mask filter f + α( f f g) = (1 + α) f α f g = f ((1 + α) e αg) image blurred image unit impulse (identity) unit impulse Gaussian Laplacian of Gaussian

44 Hybrid Images Gaussian Filter! A. Oliva, A. Torralba, P.G. Schyns, Hybrid Images, SIGGRAPH 2006 Laplacian Filter! unit impulse Gaussian Laplacian of Gaussian

45 Salvador Dali Gala Contemplating the Mediterranean Sea, which at 30 meters becomes the portrait of Abraham Lincoln, 1976

46 Band-pass filtering Gaussian Pyramid (low-pass images) Laplacian Pyramid (subband images) Created from Gaussian pyramid by subtraction

47 Laplacian Pyramid Need this! Original image How can we reconstruct (collapse) this pyramid into the original image?

48 Blending

49 Alpha Blending / Feathering I blend = αi left + (1-α)I right =

50 Affect of Window Size 1 left 1 0 right 0

51 Affect of Window Size

52 Good Window Size 1 0 Optimal Window: smooth but not ghosted

53 What is the Optimal Window? To avoid seams window = size of largest prominent feature To avoid ghosting window <= 2*size of smallest prominent feature Natural to cast this in the Fourier domain largest frequency <= 2*size of smallest frequency image frequency content should occupy one octave (power of two) FFT

54 What if the Frequency Spread is Wide FFT Idea (Burt and Adelson) Compute F left = FFT(I left ), F right = FFT(I right ) Decompose Fourier image into octaves (bands) F left = F left 1 + F left 2 + Feather corresponding octaves F left i with F right i Can compute inverse FFT and feather in spatial domain Sum feathered octave images in frequency domain Better implemented in spatial domain

55 Octaves in the Spatial Domain Lowpass Images Bandpass Images

56 Pyramid Blending Left pyramid blend Right pyramid

57 Pyramid Blending

58 laplacian level 4 laplacian level 2 laplacian level 0 left pyramid right pyramid blended pyramid

59 Blending Regions

60 Laplacian Pyramid: Blending General Approach: 1. Build Laplacian pyramids LA and LB from images A and B 2. Build a Gaussian pyramid GR from selected region R 3. Form a combined pyramid LS from LA and LB using nodes of GR as weights: LS(i,j) = GR(I,j,)*LA(I,j) + (1-GR(I,j))*LB(I,j) 4. Collapse the LS pyramid to get the final blended image

61 Horror Photo david dmartin (Boston College)

62 Results from this class (fall 2005) Chris Cameron

63 Simplification: Two-band Blending Brown & Lowe, 2003 Only use two bands: high freq. and low freq. Blends low freq. smoothly Blend high freq. with no smoothing: use binary alpha

64 High frequency (λ < 2 pixels) 2-band Laplacian Stack Blending Low frequency (λ > 2 pixels)

65 Linear Blending

66 2-band Blending

67 Da Vinci and Peripheral Vision

68 Leonardo playing with peripheral vision Livingstone, Vision and Art: The Biology of Seeing

69 Clues from Human Perception Early processing in humans filters for various orientations and scales of frequency Perceptual cues in the mid frequencies dominate perception When we see an image from far away, we are effectively subsampling it Early Visual Processing: Multi-scale edge and blob filters

70 Frequency Domain and Perception Campbell-Robson contrast sensitivity curve

71 Freq. Perception Depends on Color R G B

72 Lossy Image Compression (JPEG) Block-based Discrete Cosine Transform (DCT)

73 Using DCT in JPEG The first coefficient B(0,0) is the DC component, the average intensity The top-left coeffs represent low frequencies, the bottom right high frequencies

74 Image compression using DCT Quantize More coarsely for high frequencies (which also tend to have smaller values) Many quantized high frequency values will be zero Encode Filter responses Can decode with inverse dct Quantization table Quantized values

75 JPEG Compression Summary Subsample color by factor of 2 People have bad resolution for color Split into blocks (8x8, typically), subtract 128 For each block a. Compute DCT coefficients b. Coarsely quantize Many high frequency components will become zero c. Encode (e.g., with Huffman coding)

76 Block size in JPEG Block size small block faster correlation exists between neighboring pixels large block better compression in smooth regions It s 8x8 in standard JPEG

77 JPEG compression comparison 89k 12k

78 Denoising Gaussian Filter Additive Gaussian Noise

79 Reducing Gaussian noise Smoothing with larger standard deviations suppresses noise, but also blurs the image Source: S. Lazebnik

80 Reducing salt-and-pepper noise by Gaussian smoothing 3x3 5x5 7x7

81 Alternative idea: Median filtering A median filter operates over a window by selecting the median intensity in the window Is median filtering linear? Source: K. Grauman

82 Median filter What advantage does median filtering have over Gaussian filtering? Robustness to outliers Source: K. Grauman

83 Median filter Salt-and-pepper noise Median filtered MATLAB: medfilt2(image, [h w]) Source: M. Hebert

84 Median vs. Gaussian filtering 3x3 5x5 7x7 Gaussian Median

85

86 EXTRA SLIDES

87 A Gentle Introduction to Bilateral Filtering and its Applications Fixing the Gaussian Blur : the Bilateral Filter Sylvain Paris MIT CSAIL

88 Blur Comes from Averaging across Edges input * output * * Same Gaussian kernel everywhere.

89 Bilateral Filter [Aurich 95, Smith 97, Tomasi 98] No Averaging across Edges input * output * * The kernel shape depends on the image content.

90 Bilateral Filter Definition: an Additional Edge Term Same idea: weighted average of pixels. BF I] = new 1 W G not new new ( p q ) G ( I I ) [ p σ s σ r p q S p q I q normalization factor space weight range weight I

91 Illustration a 1D Image 1D image = line of pixels Better visualized as a plot pixel intensity pixel position

92 Gaussian Blur and Bilateral Filter Gaussian blur p q GB I = G [ ] p σ q S ( p q ) space I q space Bilateral filter [Aurich 95, Smith 97, Tomasi 98] q p space range 1 BF [ I] p = Gσ W s σ r p q S space normalization ( p q ) G ( I I ) p range q I q

93 This image cannot currently be displayed. Bilateral Filter on a Height Field 1 BF [ I] p = Gσ r W s σ p q S ( p q ) G ( I I ) p q I q p q output input reproduced from [Durand 02]

94 Space and Range Parameters BF 1 [ I] = p G r W p q S ( p q ) G ( I I ) σ s σ p q I q space σ s : spatial extent of the kernel, size of the considered neighborhood. range σ r : minimum amplitude of an edge

95 Influence of Pixels Only pixels close in space and in range are considered. space range p

96 Exploring the Parameter Space σ r = 0.1 σ r = 0.25 σ r = (Gaussian blur) input σ s = 2 σ s = 6 σ s = 18

97 Varying the Range Parameter σ r = 0.1 σ r = 0.25 σ r = (Gaussian blur) input σ s = 2 σ s = 6 σ s = 18

98 input

99 σ r = 0.1

100 σ r = 0.25

101 σ r = (Gaussian blur)

102 Varying the Space Parameter σ r = 0.1 σ r = 0.25 σ r = (Gaussian blur) input σ s = 2 σ s = 6 σ s = 18

103 input

104 σ s = 2

105 σ s = 6

106 σ s = 18