Tweet
Tweet

Slide 1

Slide 1 text

Digital Image Processing Frequency domain processing (Week 4,5) NHSM - 4th year - Fall 2024 - Prof. Mohammed Hachama [email protected] http://hachama.github.io/home/

Slide 2

Slide 2 text

Outline Fourier series and Transform Discrete Fourier Transform (DFT) Filtering in the frequency domain Sampling, FT and Shannon theorem Other transforms NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 2/28

Slide 3

Slide 3 text

Motivation • A significant part of image analysis involves the use of filters. • Fourier transform: enables replacing computationally intensive operations with faster alternatives. • 1960, FFT: fast algorithm allowing the computation of the Fourier transform (of images). NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 3/28

Slide 4

Slide 4 text

Motivation • Convert from spatial domain to frequency domain. • Perform manipulations in the frequency domain. • (Re)convert the solution back to the spatial domain. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 3/28

Slide 5

Slide 5 text

Fourier series and Transform NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 4/28

Slide 6

Slide 6 text

Periodic functions Sum of sine waves • A sine wave (or sinusoidal) f (t) = a cos(2πut + ϕ) is periodic • ∀t, f (t + T) = f (t) ; T is the period and u = 1/T is the frequency ; • a: amplitude ; ϕ: phase (ϕ = −2πus) ; s: shift NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 7

Slide 7 text

Periodic functions Sum of sine waves NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 8

Slide 8 text

Periodic functions Sum of sine waves NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 9

Slide 9 text

Periodic functions Decomposition of a T-periodic function • Harmonics = sinusoidal waves with a frequency that is a positive integer multiple of the fundamental frequency of a periodic signal NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 10

Slide 10 text

Periodic functions Decomposition of a T-periodic function Changing representation using Fourier series NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 11

Slide 11 text

Periodic functions Decomposition of a T-periodic function NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 12

Slide 12 text

Periodic functions Decomposition of a T-periodic function NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 13

Slide 13 text

Periodic functions Decomposition of a T-periodic function Image = weighted sum of sine waves NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 5/28

Slide 14

Slide 14 text

Fourier series • Decomposition of a real T-periodic function into harmonics. f (x) = a0 2 + +∞ n=1 an cos( 2π T nx) + bn sin( 2π T nx) an = 2 T T 2 − T 2 f (t) cos( 2π T nt) dt bn = 2 T T 2 − T 2 f (t) sin( 2π T nt) dt NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 6/28

Slide 15

Slide 15 text

Fourier series • Decomposition of a complex T-periodic function. f (x) = +∞ n=−∞ cn ei 2π T nx , cn = 1 T T 2 − T 2 f (t)e−i 2π T nt dt. • Fourier bases: {ei 2π T nt, n ∈ Z} is an orthonormal basis of the Hilbert space L2([−π, π]), endowed with the scalar product ⟨f |g⟩ = 1 T T 2 − T 2 f (t)¯ g(t) dt. • Convergences: in L2 norm. Pointwise convergence? NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 6/28

Slide 16

Slide 16 text

Fourier series • In image processing (or in general) • Functions (images) are not cyclic in general. • Little Use for Fourier Series. • Any L2 (or L1) function (even aperiodic) can be written as the integral of sine and cosine functions multiplied by coefficients. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 6/28

Slide 17

Slide 17 text

Fourier transform Direct and Inverse Fourier Transform • f ∈ L1(R): F(u) = +∞ −∞ f (x) e−i u x dx. F is bounded and continuous. • Inverse transform: If f , F ∈ L1(R): f (x) = 1 2π +∞ −∞ F(u) ei u x du. • Extension to L2(R) by density of L1(R) ∩ L2(R). FT generalizes FS to non-periodic functions → Continuous frequency spectrum. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 7/28

Slide 18

Slide 18 text

Fourier transform Direct and Inverse Fourier Transform • Example: f = 1[−1,+1] ∈ L2(R) F(u) = +1 −1 e−i u t dt = 2 sin u u . F is not necessarily integrable. We cannot apply the inverse transform. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 7/28

Slide 19

Slide 19 text

Fourier transform Direct and Inverse Fourier Transform • Theorem (Convolution) If f , h ∈ L1(R). Then, g = h ∗ f ∈ L1(R) and G(u) = H(u)F(u). • Theorem (Parseval and Plancherel) If f and h are two functions of L1(R) ∩ L2(R) ∞ −∞ f (t)¯ h(t) dt = 1 2π ∞ −∞ F(t)¯ H(t) dt ∞ −∞ |f (t)|2 dt = 1 2π ∞ −∞ |F(t)|2 dt NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 7/28

Slide 20

Slide 20 text

Fourier transform Direct and Inverse Fourier Transform • Properties NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 7/28

Slide 21

Slide 21 text

Fourier transform Direct and Inverse Fourier Transform • Properties NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 7/28

Slide 22

Slide 22 text

Fourier transform Direct and Inverse Fourier Transform • For f ∈ L1(R2): F(u1, u2 ) = +∞ −∞ +∞ −∞ f (x1, x2 ) e−i u x dx dy. • Inverse transform f , F ∈ L1(R2) f (x1, x2 ) = 1 4π2 +∞ −∞ +∞ −∞ F(u1, u2 ) ei u x du dv. • Parseval, Plancherel and Convolution. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 7/28

Slide 23

Slide 23 text

Discrete Fourier Transform (DFT) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 8/28

Slide 24

Slide 24 text

Discrete Fourier Transform (1) • DFT (Discrete Fourier transform): f ∈ CN F(k) = N−1 n=0 f [n] exp −i 2π N kn , k ∈ {0, 1, ..., N − 1} • We can show that (Inverse transform): f (n) = 1 N N−1 k=0 F[k] exp i 2π N kn , n ∈ {0, 1, ..., N − 1} • Circular Convolution: f ∗ h[n] = N p=0 f [p]h[n − p] , n ∈ {0, 1, ..., N − 1} NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 25

Slide 25 text

Discrete Fourier Transform (1) • DFT (Example for N = 4) • Coefficients of the matrix: F(m, n) = wmn, with w = e−j 2π N . NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 26

Slide 26 text

Discrete Fourier Transform (1) • iDFT (Example for N = 4) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 27

Slide 27 text

Discrete Fourier Transform (1) • The DFT is defined for any length N. However, N is often assumed to be a power of 2 because efficient algorithms exist for these values. • If N is not a power of 2, the signal is padded with zeros. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 28

Slide 28 text

Discrete Fourier Transform (1) • The family {ek [n] = exp 2π N ikn , 0 ≤ k ≤ N} is an orthogonal basis of the space of N-periodic signals, endowed with the scalar product ⟨f , h⟩ = N−1 k=0 f [k]¯ h[k] • Parseval: ∥f ∥2= N−1 k=0 |f [k]|2 = 1 N N−1 k=0 |F[k]|2 • Convolution: If g = f ∗ h, then : G[k] = F[k]H[k] NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 29

Slide 29 text

Discrete Fourier Transform (1) • Magnitude or spectrum (F = R + iI): |F(u)|= R(u)2 + I(u)2 1/2 • Power spectrum, spectral density (Energy measurement): |F(u)|2= R(u)2 + I(u)2 • Angle or phase: Φ(u) = arctan( I(u) R(u) ) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 30

Slide 30 text

Discrete Fourier Transform (1) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 31

Slide 31 text

Discrete Fourier Transform (1) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 32

Slide 32 text

Discrete Fourier Transform (1) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 33

Slide 33 text

Discrete Fourier Transform (1) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 34

Slide 34 text

Discrete Fourier Transform (1) • DFT in 2D (DFT2): F(u, v) = M−1 x=0 N−1 y=0 f (x, y) e−i 2π(ux/M+vy/N) with u = 0, ..., M − 1; v = 0, ..., N − 1. • Inverse transform: f (x, y) = 1 MN M−1 u=0 N−1 v=0 F(u, v) ei 2π(ux/M+vy/N) with x = 0, ..., M − 1; y = 0, ..., N − 1. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 35

Slide 35 text

Discrete Fourier Transform (1) • The centered format of the 2D Discrete Fourier Transform (DFT2), with N even: X(k, l) = M 2 −1 m=− M 2 N 2 −1 n=− N 2 x(m, n) e−i 2π(km/M+ln/N) with k = 0, ..., M − 1; l = 0, ..., N − 1. x(m, n) = 1 MN M 2 −1 k=− M 2 N 2 −1 l=− N 2 X(k, l) ei 2π(km/M+ln/N) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 36

Slide 36 text

Discrete Fourier Transform (1) Bases Bu,v (x, y) = 1 √ M √ N e2πj(ux N + vy M ), u = 1, ..., N − 1, v = 1, ..., M − 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 37

Slide 37 text

Discrete Fourier Transform (1) Bases Bu,v (x, y) = 1 √ M √ N e2πj(ux N + vy M ), u = 1, ..., N − 1, v = 1, ..., M − 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 38

Slide 38 text

Discrete Fourier Transform (1) Bases Bu,v (x, y) = 1 √ M √ N e2πj(ux N + vy M ), u = 1, ..., N − 1, v = 1, ..., M − 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 39

Slide 39 text

Discrete Fourier Transform (1) Bases Bu,v (x, y) = 1 √ M √ N e2πj(ux N + vy M ), u = 1, ..., N − 1, v = 1, ..., M − 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 40

Slide 40 text

Discrete Fourier Transform (1) Bases Bu,v (x, y) = 1 √ M √ N e2πj(ux N + vy M ), u = 1, ..., N − 1, v = 1, ..., M − 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 41

Slide 41 text

Discrete Fourier Transform (1) Bases Bu,v (x, y) = 1 √ M √ N e2πj(ux N + vy M ), u = 1, ..., N − 1, v = 1, ..., M − 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 42

Slide 42 text

Discrete Fourier Transform (1) • Each term of F(u, v) is a function of ALL the values of f (x, y) weighted by the exponent. • The Fourier transform presents • Average at the origin (DC component) • Low frequencies • High frequencies NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 43

Slide 43 text

Discrete Fourier Transform (1) • The DFT is conjugate symmetric F(u, v) = F(−u, −v) • If f (x, y) is real |F(u, v)|= |F(−u, −v)| The spectrum is symmetric! • In many Fourier transforms, there are residual vertical and horizontal lines passing through the origin. They come from the edge effects of the discrete transform. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 44

Slide 44 text

Discrete Fourier Transform (1) • Discrete Convolution f ∗ h(x, y) = 1 MN M m=0 N n=0 f (m, n)h(x − m, y − n) • Correspondence between spatial and spectral filters. • Convolution and product for direct and inverse transforms. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 45

Slide 45 text

Discrete Fourier Transform (1) • Properties of the DFT2 • Linearity: ax1 (m, n) + bx2 (m, n) ↔ aX1 (k, l) + bX2 (k, l) • Periodicity x(m, n) = x(m + aM, n + bN), for all m, n X(k, l) = X(k + aM, l + bN), for all k, l • Circular Shift of the image x(m − m0, n − n0 ) ↔ X(k, l)e−j 2π N (km0+ln0) • Circular Shift of the spectrum x(m, n)ej 2π N (k0m+l0n) ↔ X(k − k0, l − l0 ) The centered spectrum is obtained pour odd N and k0 = l0 = N/2. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 46

Slide 46 text

Discrete Fourier Transform (1) • Properties of the DFT2 • Circular convolution in the spatial domain x(m, n) ∗ h(m, n) ↔ X(k, l) H(k, l) y(m, n) = N−1 p=0 N−1 q=0 x(p, q)h(m − p, n − q), m, n = 0, 1, . . . , N − 1 • Circular convolution in the frequency domain x(m, n) h(m, n) ↔ 1 N2 N−1 p=0 N−1 q=0 X(p, q)H(m − p, n − q) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 47

Slide 47 text

Discrete Fourier Transform (1) Python code % Import libraries import cv2 as cv import numpy as np % Read data im = cv.imread(’leopard.jpg’,0) % Compute the FFT of the image fft = np.fft.fft2(image) % Shift the zero frequency component to the center fft˙shifted = np.fft.fftshift(fft) % Compute the magnitude (spectrum) and phase magnitude˙spectrum = np.log(np.abs(fft˙shifted) + 1) # Add 1 to avoid log(0) phase˙spectrum = np.angle(fft˙shifted) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 9/28

Slide 48

Slide 48 text

Discrete Fourier Transform (2) Interpretation of the spectrum • Shifting the zero frequency NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 49

Slide 49 text

Discrete Fourier Transform (2) Interpretation of the spectrum NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 50

Slide 50 text

Discrete Fourier Transform (2) Interpretation of the spectrum • A small object in space has a large frequency extent and vice-versa. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 51

Slide 51 text

Discrete Fourier Transform (2) Interpretation of the spectrum • Interpretation of the spectrum NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 52

Slide 52 text

Discrete Fourier Transform (2) Interpretation of the spectrum • The Fourier Transform of an Edge NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 53

Slide 53 text

Discrete Fourier Transform (2) Interpretation of the spectrum • The Fourier Transform of an Edge NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 54

Slide 54 text

Discrete Fourier Transform (2) Interpretation of the spectrum • The Fourier Transform of a bar NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 10/28

Slide 55

Slide 55 text

Discrete Fourier Transform (3) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 56

Slide 56 text

Discrete Fourier Transform (3) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 57

Slide 57 text

Discrete Fourier Transform (3) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 58

Slide 58 text

Discrete Fourier Transform (3) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 59

Slide 59 text

Discrete Fourier Transform (3) Examples NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 60

Slide 60 text

Discrete Fourier Transform (3) Reconstruction NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 61

Slide 61 text

Discrete Fourier Transform (3) Numerical examples • Calculate the DFT2, iDFT2, DFT row-wise and then column-wise, DFT column-wise and then row-wise. We assume that the origin (0, 0) is at the top left corner. 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 , 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 62

Slide 62 text

Discrete Fourier Transform (3) Numerical examples • Calculate the DFT2 column-wise and then row-wise. 1 2 3 1 −2 3 1 4 1 1 2 2 3 1 2 4 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 63

Slide 63 text

Discrete Fourier Transform (3) Numerical examples • Calculate the 1D DFT row-wise, column-wise, 2D DFT, and centered 2D DFT NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 64

Slide 64 text

Discrete Fourier Transform (3) Numerical examples • Example: DFT1 row-wise • Example: DFT1 column-wise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 65

Slide 65 text

Discrete Fourier Transform (3) Numerical examples • Example: DFT2 • Example: centered DFT2 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 66

Slide 66 text

Discrete Fourier Transform (3) Numerical examples • Example: DFT2 - magnitude NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 67

Slide 67 text

Discrete Fourier Transform (3) Numerical examples • Example: DFT2 - magnitude - logarithmic scale NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 11/28

Slide 68

Slide 68 text

Filtering in the frequency domain NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 12/28

Slide 69

Slide 69 text

Filtering in the frequency domain (1) Convolution ≈ multiplication NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 13/28

Slide 70

Slide 70 text

Filtering in the frequency domain (1) Example: Laplacian Filter x(m, n) = 1 −1 3 2 2 1 2 4 1 −1 2 −2 3 1 2 2 , h(m, n) = 0 −1 0 −1 5 −1 0 −1 0 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 13/28

Slide 71

Slide 71 text

Filtering in the frequency domain (1) Example: Laplacian Filter • Border replication x = 1 −1 3 2 2 1 2 4 1 −1 2 −2 3 1 2 2 , X = 1 1 −1 3 2 2 0 0 1 1 −1 3 2 2 0 0 2 2 1 2 4 4 0 0 1 1 −1 2 −2 −2 0 0 3 3 1 2 2 2 0 0 3 3 1 2 2 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 13/28

Slide 72

Slide 72 text

Filtering in the frequency domain (1) Example: Laplacian Filter • Zero-padding h = 0 −1 0 −1 5 −1 0 −1 0 , H = 5 −1 0 0 0 0 0 −1 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 13/28

Slide 73

Slide 73 text

Filtering in the frequency domain (1) Example: Laplacian Filter • Result y(m, n) = 2 −9 9 −1 5 3 0 14 0 −10 9 −16 7 0 3 6 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 13/28

Slide 74

Slide 74 text

Filtering in the frequency domain (2) Frequency domain filters • Band-Stop Filtering (Notch Filtering): Remove specific frequencies or bands ; Often used to eliminate periodic noise. • Low-pass filters: Eliminate high frequencies ; Remove details. • High-pass filters: Eliminate low frequencies ; Preserve the outline of objects. • Band-Pass Filtering: Retain frequencies within a specific range (band) ; Remove frequencies outside this range. • Homomorphic Filtering: Enhances the contrast of an image by manipulating both high and low frequencies. • Unsharp Masking: To sharpen an image by subtracting a blurred version from the original. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 14/28

Slide 75

Slide 75 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Band-stop filtering removes frequencies within a specific range, often used to eliminate periodic noise or interference. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 76

Slide 76 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Remove the DC component: subtract the average intensity H(u, v) = 0 (u, v) = (0, 0) 1 Elsewhere • Emphasize the high-frequency components (edges and fine textures). • Sometimes, normalize the image for further processing steps. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 77

Slide 77 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Periodic noise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 78

Slide 78 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Periodic noise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 79

Slide 79 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Periodic noise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 80

Slide 80 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Periodic noise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 81

Slide 81 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Periodic noise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 82

Slide 82 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Moir´ e Patterns: wavy lines (curves), grids, stripes, ripples, or repetitive patterns. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 83

Slide 83 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Moir´ e Patterns: wavy lines (curves), grids, stripes, ripples, or repetitive patterns. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 84

Slide 84 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Moir´ e Patterns: wavy lines (curves), grids, stripes, ripples, or repetitive patterns. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 85

Slide 85 text

Filtering in the frequency domain (2) Band-Stop Filtering (Notch Filtering) • Moir´ e Patterns: wavy lines (curves), grids, stripes, ripples, or repetitive patterns. • Origins • Interference patterns created when two regular grids or patterns overlap. • Scanner Artifacts: Regular grid structures in scanners. • Optical Issues: Lens imperfections or optical filters. • Moir´ e patterns are periodic and can be removed using frequency domain techniques. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 15/28

Slide 86

Slide 86 text

Filtering in the frequency domain (2) Low-pass filter • Ideal low-pass filter: Cuts off all high frequencies beyond a distance D0 (cutoff frequency) from the center. H(u, v) = 1 (u − u0 )2 + (v − v0 )2 < D0 0 Elsewhere NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 87

Slide 87 text

Filtering in the frequency domain (2) Low-pass filter • Ideal low-pass filter: Cuts off all high frequencies beyond a distance D0 (cutoff frequency) from the center. • Removes high-frequency components and retains low-frequency components → results in a blurred image, reducing detail and noise. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 88

Slide 88 text

Filtering in the frequency domain (2) Low-pass filter • Ideal low-pass filter: Cuts off all high frequencies beyond a distance D0 (cutoff frequency) from the center. H(u, v) = 1 (u − u0 )2 + (v − v0 )2 < D0 1/2 Elsewhere NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 89

Slide 89 text

Filtering in the frequency domain (2) Low-pass filter • Ideal low-pass filter: Reverberation phenomenon NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 90

Slide 90 text

Filtering in the frequency domain (2) Low-pass filter • Ideal low-pass filter: Reverberation phenomenon NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 91

Slide 91 text

Filtering in the frequency domain (2) Low-pass filter • Butterworth low-pass filter: Gradually cuts off high frequencies based on the selection of D0 and the exponent n H(u, v) = 1 1 + Dist((u,v);(u0,v0)) D0 2n . NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 92

Slide 92 text

Filtering in the frequency domain (2) Low-pass filter • Butterworth low-pass filter: Gradually cuts off high frequencies based on the selection of D0 and the exponent n NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 93

Slide 93 text

Filtering in the frequency domain (2) Low-pass filter • Butterworth low-pass filter: Gradually cuts off high frequencies based on the selection of D0 and the exponent n NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 94

Slide 94 text

Filtering in the frequency domain (2) Low-pass filter • Butterworth low-pass filter: Gradually cuts off high frequencies based on the selection of D0 and the exponent n NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 95

Slide 95 text

Filtering in the frequency domain (2) Low-pass filter • Gaussian low-pass filter: Gradually cuts off high frequencies based on the selection of D0 H(u, v) = e − D(u,v)2 2D2 0 . NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 96

Slide 96 text

Filtering in the frequency domain (2) Low-pass filter • Gaussian low-pass filter: Gradually cuts off high frequencies based on the selection of D0 H(u, v) = e − D(u,v)2 2D2 0 . • Less aggressive than the ideal filter or Butterworth filter. • Less control over the precise selection of D0 . • But provides protection against reverberation! • Matches the human model. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 16/28

Slide 97

Slide 97 text

Filtering in the frequency domain (2) High-pass filter • Exactly the opposite function of low-pass filters Hph = 1 − Hpb • Ideal, Butterworth, Gaussian NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 17/28

Slide 98

Slide 98 text

Filtering in the frequency domain (2) High-pass filter • Exactly the opposite function of low-pass filters Hph = 1 − Hpb • Ideal, Butterworth, Gaussian • Ideal High-pass 0 — 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 17/28

Slide 99

Slide 99 text

Filtering in the frequency domain (2) High-pass filter • Exactly the opposite function of low-pass filters Hph = 1 − Hpb • Ideal, Butterworth, Gaussian • Ideal High-pass 0 — 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 17/28

Slide 100

Slide 100 text

Filtering in the frequency domain (2) High-pass filter • Exactly the opposite function of low-pass filters Hph = 1 − Hpb • Ideal, Butterworth, Gaussian • Ideal High-pass 1/2 — 1 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 17/28

Slide 101

Slide 101 text

Filtering in the frequency domain (2) High-pass filter • Exactly the opposite function of low-pass filters Hph = 1 − Hpb • Ideal, Butterworth, Gaussian • Gaussian High-pass NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 17/28

Slide 102

Slide 102 text

Filtering in the frequency domain (2) Band-pass filter Retain frequencies within a specific range and remove those outside this range (useful for isolating certain details or patterns). NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 18/28

Slide 103

Slide 103 text

Filtering in the frequency domain (2) Homomorphic filter Enhance the contrast of an image by modifying both high and low frequencies (useful for dynamic range compression and contrast enhancement). NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 19/28

Slide 104

Slide 104 text

Filtering in the frequency domain (2) Homomorphic filter 1. Take the logarithm of the image. 2. Apply a high-pass filter to attenuate low-frequency components and boost high-frequency components. H(u, v) = (γH − γL ) · 1 − e−c·((u−u0)2+(v−v0)2) + γL where γH, γL : Gain for high frequencies (reflectance) and low frequencies (illumination), c controls the filter’s sharpness, 3. Take the inverse Fourier transform and exponentiate the result. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 19/28

Slide 105

Slide 105 text

Filtering in the frequency domain (2) Homomorphic filter • ⇝ A gradual change in illumination in the left image (original) has been corrected in the image on the right (filtered image). NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 19/28

Slide 106

Slide 106 text

Filtering in the frequency domain (2) Homomorphic filter • Enhancing images with uneven lighting, • Improving contrast in images with shadows or bright areas, • Biomedical image enhancement, particularly for visualizing tissue details in medical scans. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 19/28

Slide 107

Slide 107 text

Filtering in the frequency domain (2) Spectral filtering – Image sharpening • Image sharpening = Re-enforce edges • How: Keep low frequencies + Amplify high frequencies • y = x + αDx, D: derivative filter, α > 0 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 20/28

Slide 108

Slide 108 text

Filtering in the frequency domain (2) Spectral filtering – Image sharpening • Drawback: Amplify noise NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 20/28

Slide 109

Slide 109 text

Sampling, FT and Shannon theorem NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 21/28

Slide 110

Slide 110 text

Sampling and FT Aliasing • Original image NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 111

Slide 111 text

Sampling and FT Aliasing • Sampled image: Aliasing effect NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 112

Slide 112 text

Sampling and FT Aliasing • Sampling a signal: Aliasing effect • Undersampling a signal: different frequencies. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 113

Slide 113 text

Sampling and FT Aliasing • Sampling a signal: Aliasing effect • Undersampling a signal: different frequencies. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 114

Slide 114 text

Sampling and FT Sampling and reconstruction NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 115

Slide 115 text

Sampling and FT Sampling and reconstruction ˆ fd (w) = ∞ n=−∞ f (ns)e−insw fd (t) = ∞ n=−∞ f (ns)δ(t − ns) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 116

Slide 116 text

Sampling and FT Sampling and reconstruction ˆ fd (w) = ∞ n=−∞ f (ns)e−insw fd (t) = ∞ n=−∞ f (ns)δ(t − ns) ˆ fd (w) = 1 s ∞ k=−∞ ˆ f (w − 2kπ s ) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 117

Slide 117 text

Sampling and FT Sampling and reconstruction NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 118

Slide 118 text

Sampling and FT Sampling and reconstruction • Shannon Sampling Theorem: If Supp(ˆ f ) = −π s , π s , then: f (t) = ∞ n=−∞ f (ns)Φs (t − ns), Φs (t) = sin(πt/s) πt/s . NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 119

Slide 119 text

Sampling and FT Sampling and reconstruction NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 120

Slide 120 text

Sampling and FT Statement of Shannon’s Theorem A band-limited continuous signal can be completely represented by its samples and fully reconstructed if it is sampled at a rate greater than twice its highest frequency: fs > 2fmax where: • fs is the sampling frequency. • fmax is the maximum frequency of the signal. This rate is known as the Nyquist Rate. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 121

Slide 121 text

Sampling and FT Shannon’s Theorem and Aliasing • Sampling below the Nyquist rate leads to aliasing. • Correct sampling ensures that the original signal can be perfectly reconstructed from its samples. • Anti-Aliasing Techniques • Technique 1: • Technique 2: NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 122

Slide 122 text

Sampling and FT Shannon’s Theorem and Aliasing • Sampling below the Nyquist rate leads to aliasing. • Correct sampling ensures that the original signal can be perfectly reconstructed from its samples. • Anti-Aliasing Techniques • Technique 1: Higher Sampling Rates: Ensuring sampling is above the Nyquist rate. • Technique 2: NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 123

Slide 123 text

Sampling and FT Shannon’s Theorem and Aliasing • Sampling below the Nyquist rate leads to aliasing. • Correct sampling ensures that the original signal can be perfectly reconstructed from its samples. • Anti-Aliasing Techniques • Technique 1: Higher Sampling Rates: Ensuring sampling is above the Nyquist rate. • Technique 2: Filtering: Applying a low-pass filter before sampling to remove high-frequency components. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 124

Slide 124 text

Sampling and FT Anti-Aliasing filter Input Reconst. without AA Reconst. with AA NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 125

Slide 125 text

Sampling and FT Anti-Aliasing filter Input Reconst. without AA Reconst. with AA NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 126

Slide 126 text

Sampling and FT Anti-Aliasing filter Input Reconst. without AA Reconst. with AA NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 127

Slide 127 text

Sampling and FT Anti-Aliasing filter Input Reconst. without AA Reconst. with AA NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 22/28

Slide 128

Slide 128 text

Other transforms NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 23/28

Slide 129

Slide 129 text

Discrete Cosine Transform (DCT) What is DCT? • The Discrete Cosine Transform (DCT) transforms a signal or image from the spatial domain to the frequency domain. • The DCT is similar to the Fourier Transform, but it uses only cosine functions and is real function, making it more efficient for image compression. • DCT is widely used in image processing, especially in compression techniques like JPEG, due to its ability to concentrate energy in the lower frequencies. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 130

Slide 130 text

Discrete Cosine Transform (DCT) Formulas • 1D DCT formula for a sequence x[n] of length N: X[k] = α(k) N−1 n=0 x[n] cos π N n + 1 2 k , k = 0, 1, . . . , N − 1 α(k) =    1 N for k = 0 2 N for k = 1, 2, . . . , N − 1 • Inverse 1D DCT x[n] = N−1 k=0 α(k)X[k] cos π N n + 1 2 k , n = 0, 1, . . . , N − 1. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 131

Slide 131 text

Discrete Cosine Transform (DCT) Formulas • The 2D DCT of an M × N image x[m, n] X[k, l] = α(k)β(l) M−1 m=0 N−1 n=0 x[m, n] cos π M m + 1 2 k cos π N n + 1 2 l where k = 0, . . . , M − 1, l = 0, . . . , N − 1, and the normalization factors: α(k) =    1 M for k = 0 2 M for k = 1, . . . , M − 1 , β(l) =    1 N for l = 0 2 N for l = 1, . . . , N − 1 • The 2D DCT of an (8 × 8) image x[m, n] X(k, l) = 1 4 7 m=0 7 n=0 x(m, n) · cos (2m + 1)kπ 16 · cos (2n + 1)lπ 16 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 132

Slide 132 text

Discrete Cosine Transform (DCT) Formulas • Inverse 2D DCT x[m, n] = M−1 k=0 N−1 l=0 α(k)α(l)X[k, l] cos π M m + 1 2 k cos π N n + 1 2 l where m = 0, 1, . . . , M − 1 and n = 0, 1, . . . , N − 1. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 133

Slide 133 text

Discrete Cosine Transform (DCT) DCT basis • DCT basis are cosine functions that represent different frequencies. • Low-frequency components (coarse features) and high-frequency components (fine details). NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 134

Slide 134 text

Discrete Cosine Transform (DCT) DCT and Image Compression • After transformation to the frequency domain, most of the energy is concentrated in the lower frequencies. • The high-frequency components, which correspond to small details, can be quantized more coarsely or discarded, leading to compression. • A simple compression method: apply a threshold on the DCT coefficients. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 135

Slide 135 text

Discrete Cosine Transform (DCT) DCT and Image Compression • DCT compression: Apply a threshold on the DCT coefficients Original (154401 coeffs.) DCT (5214 coeffs.) Compressed 3.38% NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 136

Slide 136 text

Discrete Cosine Transform (DCT) DCT and Image Compression • DCT compression: Apply a threshold on the DCT coefficients Compression rate: 3.38% NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 137

Slide 137 text

Discrete Cosine Transform (DCT) DCT and Image Compression • DCT compression: Apply a threshold on the DCT coefficients Compression rate: 20% NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 138

Slide 138 text

Discrete Cosine Transform (DCT) DCT and Image Compression • DCT in JPEG Compression 1. Convert the image to the YCbCr color space. 2. Split the image into 8 × 8 blocks. 3. Apply the 2D DCT to each block. 4. Quantize the DCT coefficients. 5. Encode the quantized coefficients using entropy coding. →Chapter 4 NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 24/28

Slide 139

Slide 139 text

Wavelets What are Wavelets? • A wavelet is a small wave, localized in time • Decompose data into different frequency components. • Unlike Fourier transform, which uses sinusoids, wavelet transform uses functions that are localized in both time and frequency. • Effective for analyzing non-stationary signals and images. • Wavelets provide a multi-resolution analysis of images. • Many applications: compression (like JPEG 2000), denoising, feature extraction and texture analysis. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 25/28

Slide 140

Slide 140 text

Wavelet Transforms Continuous Wavelet Transforms • 1D Continuous Wavelet Transform (CWT) Wx (a, b) = ∞ −∞ x(t) 1 |a| ψ t − b a dt where a, b represent scaling and translation and ψ is a wavelet mother satisfying +∞ −∞ ψ(t)dt = 0, ||ψ||= 1. • {ψa,b (t) = 1 √ |a| ψ t−b a }a,b is a dictionary (overcomplete and redundant). Reconstruction, Bases? CWT = a magnifying glass. You adjust the scale to zoom in on specific parts of the signal, and you translate to examine different time instances. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 141

Slide 141 text

Wavelet Transforms Continuous Wavelet Transforms • 1D Continuous Wavelet Transform (CWT) Wx (a, b) = ∞ −∞ x(t) 1 |a| ψ t − b a dt CWT: Uses the mother wavelet to analyze the signal continuously across different scales and positions. • Examples • Haar Wavelet • Daubechies Wavelet • Symlets • Coiflets • ... NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 142

Slide 142 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) • 1D Wavelet Transform: decomposes a signal into approximation and detail coefficients. A1 [n] = k x[k] · g[2n − k], D1 [n] = k x[k] · h[2n − k] • g[n] is the low-pass filter, producing the approximation coefficients. • h[n] is the high-pass filter, producing the detail coefficients. • The subsampling by 2 (represented by 2n) reduces the resolution, capturing the signal at a coarser level. DWT: Uses both the mother wavelet (for details) and the father wavelet (for approximations) to decompose and reconstruct the signal in a multi-resolution framework. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 143

Slide 143 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) • Multiresolution Analysis (MRA) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 144

Slide 144 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) • 2D Wavelet Transform: applied separately to the rows and the columns. • LL: Low-pass filtering in both dimensions, capturing the approximation. • LH: Low-pass filtering along the rows and high-pass filtering along the columns, capturing horizontal details. • HL: High-pass filtering along the rows and low-pass filtering along the columns, capturing vertical details. • HH: High-pass filtering in both dimensions, capturing diagonal details. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 145

Slide 145 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) • 2D Wavelet Transform: applied separately to the rows and the columns. LL = M−1 x=0 N−1 y=0 f (x, y) · g(2x − u) · g(2y − v) LH = M−1 x=0 N−1 y=0 f (x, y) · g(2x − u) · h(2y − v) HL = M−1 x=0 N−1 y=0 f (x, y) · h(2x − u) · g(2y − v) HH = M−1 x=0 N−1 y=0 f (x, y) · h(2x − u) · h(2y − v) • g and h are the low and high-pass filters, respectively. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 146

Slide 146 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 147

Slide 147 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 148

Slide 148 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 149

Slide 149 text

Wavelet Transforms Discrete Wavelet Transforms (DWT) • Haar Wavelet • Simplest wavelet, piecewise constant. • Scaling function (father wavelet) and wavelet function (mother wavelet) defined as: ϕ(t) = 1 0 ≤ t < 1, 0 otherwise. ; ψ(t) =      1 0 ≤ t < 0.5, −1 0.5 ≤ t < 1, 0 otherwise. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 26/28

Slide 150

Slide 150 text

Wavelets applications Image Compression • Wavelet transform is used in image compression standards like JPEG 2000. • It provides a multi-resolution representation, allowing for progressive transmission and high compression ratios. • Simple method 1. Decompose the image using the Discrete Wavelet Transform (DWT). 2. Threshold the coefficients, keeping only the significant ones. 3. Reconstruct the compressed image using the Inverse Discrete Wavelet Transform (IDWT). NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 27/28

Slide 151

Slide 151 text

Wavelets applications Image Denoising with Wavelets 1. Decompose the noisy signal using the DWT. 2. Threshold the detail coefficients to remove the noise. • Hard thresholding sets coefficients below the threshold to zero. • Soft thresholding reduces coefficients by the threshold value. 3. Reconstruct the denoised signal using the IDWT. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 27/28

Slide 152

Slide 152 text

Wavelets applications Edge Detection using Wavelets • Wavelet transform can be used to detect edges in images. • The high-pass components (LH, HL, HH) contain the edge information. • By analyzing these components, we can detect edges at multiple scales. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 27/28

Slide 153

Slide 153 text

Assignment 3 Write a python code that depends on the OpenCV library and that performs all the methods of this lecture: 1. Fourier transform: illustrate the properties of the FT. 2. DFT and DFT2 and their inverses: compute and visualize some examples. 3. Illustrate the reconstruction from DFT and DFT2. 4. Frequency domain filters and their applications. 5. Sampling and FT: illustrate the aliasing, use of anti-aliasing filter. 6. DCT1 and DCT2 and application to denoising and compression. 7. DWT1 and DWT2 and application to denoising, compression and edge detection. Simulations should be conducted to evaluate (qualitative and quantitative) and compare the methods with various parameters. NHSM - 4th year: Digital Image Processing - Frequency Processing (Week 4,5) - M. Hachama ([email protected]) 28/28