Filtering

Basis beeldverwerking (8D040)

dr. Andrea FusterProf.dr. Bart ter Haar Romenydr. Anna VilanovaProf.dr.ir. Marcel Breeuwer

Filtering

Contents

• Sharpening Spatial Filters• 1st order derivatives• 2nd order derivatives• Laplacian • Gaussian derivatives• Laplacian of Gaussian (LoG)• Unsharp masking

2

Sharpening spatial filters

• Image derivatives (1st and 2nd order)• Define derivatives in terms of differences for the

discrete domain• How to define such differences?

3

1st order derivatives

• Some requirements (1st order):• Zero in areas of constant intensity• Nonzero at beginning of intensity step or ramp• Nonzero along ramps

4

1st order derivatives

5

2nd order derivatives

• Requirements (2nd order)• Zero in constant areas• Nonzero at beginning and end of intensity step or ramp• Zero along ramps of constant slope

6

2nd order derivatives

7

Image Derivatives

• 1st order

• 2nd order

8

1-1

1 1-2

9

1st order2nd order

Zero crossing, locating edges

• Edges are ramp transitions in intensity• 1st order derivative gives thick edges• 2nd order derivative gives double thin edge with zeros in

between

• 2nd order derivatives enhance fine detail much better

10

11

2nd order

Zero crossing, locating edges

1st order

Filters related to first derivatives

• Recall: Prewitt filter, Sober filter (lecture 2 – 01/05)

12

Laplacian – second derivative

• Enhances edges• Definition

13

Laplacian

14

Adding diagonal derivationOpposite sign for second order derivative

Laplacian

• Note: Laplacian filtering results in + and – pixel values

• Scale for image display • So: take absolute value or positive values

15

Line Detector

16

*

Laplacian Positive values Laplacian(figure

10.5 book)

Image sharpening - example

17

4-connected Laplacian8-connected LaplacianEnhanced + Laplacian x5Enhanced + Laplacian x6Enhanced + Laplacian x8Better sharpening with 8-connected Laplacian(see figure 3.38 (d)-(e) book)

C=+1 or -1

Filtering in frequency domain

• Basic steps:− image f(x,y) − Fourier transform F(u,v)− filter H(u,v) − H(u,v)F(u,v)− inverse Fourier transform − filtered image g(x,y)

18

Laplacian in the Fourier domain

• Spatial

• Fourier domain

19

Blur first, take derivative later

• Smoothing is a good idea to avoid enhancement of noise. Common smoothing kernel is a Gaussian.

2 2

22x y

G e

Scale of blurring

Gaussian Derivative

• Taking the derivative after blurring gives image g

*( )g D G f

2 2

22x y

G e

Gaussian Derivative

• We can build a single kernel for both convolutions

( * )g D G f 2 2

222*

x y

xxD G e

2 2

222*

x y

yyD G e

Use the associative property of the convolution

Laplacian of Gaussian (LoG)

23

LoG a.k.a. Mexican Hat

LoG applied to building

24

Sharpening with LoG

25sharpening with LoG sharpening

with Laplacian

Unsharp Masking / Highboost Filtering

• Subtraction of unsharp (smoothed) version of image from the original image.

• Blur the original image• Subtract the blurred image from the original

(results in image called mask)• Add the mask to the original

26

• Let denote the blurred image• Obtain the mask

• Add weighted portion of mask to original image

27

• If• Unsharp masking

• If• Highboost filtering

28

input blurred unsharp mask u.m. result h.f. result

(see also figure 3.40 book)

Unsharp masking

• Simple and often used sharpening method• Poor result in the presence of noise – LoG performs

better in this case

29