texture filtering fourier transform and

Center for Image Processing

Fourier Transform andTexture Filtering

Lucas J. van Vlietwww.ph.tn.tudelft.nl/~lucas

Center for Image ProcessingCenter for Image Processing

CBP course: Fourier & Texture 2


Image Analysis Paradigm

Texture filtering

analysis

segmentation

Imagerestoration

Imageenhancement

sensorImageformation scene

pre-processing

classification

Texture properties:scale

orientationLocal variance

regularity




Texture properties

� Texture is domain property rather than a pixel property.

� Fourier analysis offers a powerful tool to analyze and describe domain properties

� Properties� Local power = local variance� Scale (from fine detail to course structures)� Anisotropy (orientation preference)




Discrete Fourier Transform

� Each image can be decomposed in weighted sum of complex exponentials (sines and cosines) of frequency fand angle φ. (or two frequency components u and v)

( ) ( )( )

( ) ( )( )

1 1 2

0 0

1 1 2

0 0

1, ,

1, ,

N N j ux vyN

u v

N N j ux vyN

u v

g x y G u v eN

G u v g x y eN

π

π

− −+

= =

− −− +

= =

=

=

∑∑

∑∑

� For real-valued images: ( ){ } ( ){ }

( ){ } ( ){ }

Ev , Re ,

Od , Im ,

F

F

g x y G u v

g x y j G u v

←→

←→

image sizeNxN




Getting used to Fourier (2)

An image is a weighted sum of cos (even) and sin (odd) images.

a+jb

Fourier domain with complex amplitude: a+jb




Getting used to Fourier (1)

� Graphite surface by Scanning Tunneling Microscopy� Atomic structure of graphite shows a hexagonal surface

column

row

0

0N-1

N-1

(c,r) = (½N,½N)(u,v) = (0,0)

x

y

u

v

(c,r) = (0,0)(u,v) = (-½N,-½N)

(c,r) = (N-1,N-1)(u,v) = (½N-1, ½N-1)




Superposition� Fourier spectrum

= + + +

= + + +

F F F F F




Two domains

� The coordinate systems (x,y) and (u,v) have their origin at the position (column,row) = (½N, ½N)

u

v

x

y

column

row

column

row

0

0N-1

N-1

(u,v) = (0,0)




Local variance filter� Recipe: local variance filter(filter size = n)

� Compute the local mean (blurring filter of size n)� Subtract (if necessary) the local mean.� Compute the square of each pixel value.� Suppress the “double” response by local averaging (blurring filter

of size n).� Local variance is a measure for the local squared-contrast.




Segmenting textured objects� Simple thresholding fails!� The objects differ in texture from the background� Local variance followed by a (morphological) closing yields …




Scaling: local vs global

� Problem: Choosing the proper scale is an important, but tedious task.� Scale too small: Texture characteristics are missed

which yields an incomplete data description.� Scale too large: Confusion (mixing) of adjacent texture

characteristics, lack of localization, and blindness for small and subtle textured regions.

� Solution: Multi-scale analysis.� Analyze the image as function of scale: from fine detail

to course “image-filling” objects.

( ) ( )( ) ( )

2 22 2 22 2

22 22

12

, ,u vx y N N

Fg x y e G u v eπ π

σ

σ

πσ

− + + −

= ←→ =




Multi-Scale

Series of images filtered of decreasing scale: Scale-space� Sample the scales logarithmically using filters of size = basescale

yields n scales per octave

Input imagescale 0

scale 1

scale 2

scale 3

scale 4

scale 5

var (scale 1)

var (scale 2)

var (scale 3)

var (scale 4)

var (scale 5)

Local variance between scales n and n-1.

{ }11 13212 ,2 ,2 ,...,2 nbase ∈

Scale difference ≈ scale derivative

Scal

e sp

ace

Scal

e sp

ace




Scale-spaces

� Morphological scale-space: Use openings (closings)� Gaussian scale-space: Use Gaussian filters

Increasing scales

Fourier domain with “footprints” of Gaussianfilters of increasing scale

Filter size is inversely proportional to “footprint” in Fourier domain

12

345




Chirp exampleSc

ale-

spac

e

Scale derivative Spatial variance

Scal

e-sp

ace




Gabor: scale and orientation

� Problem: Crossing and touching of oriented patterns.� Isotropic scale cannot discriminate based on orientation.

� Solution: Multi-orientation analysis� Orientation selective filtering

� Multi-scale and multi-orientation analysis by Gabor filtering.� Orientation selectivity

( ) ( )2 2

2 222

1, 2

,x y

i iNi i

j u x v yu vg x y e e

π

σ

πσ

+−

+=

Gaussian window of size σSelected frequency (ui,vi)




Gabor filter bank

� Linear sampling of orientation� Logarithmic sampling of scale� (Remember: reciprocal relation of size in space and freq.

A single input imageyields many output images.

Question: Which output image(s) discriminate between textures?

( )

( )( ) ( )( )

( )

2 2 22 2

2, ,,

u u v vi iN N

i iu v

i i

G u v eG u u v v

π πσ

− − + −

=

= − −(ui,vi) σ−1




Gabor filtering

� Each Gabor filter selects a (Gaussian-shaped) region of the Fourier spectrum of the image.

� A Gabor filter can be implemented by convolution!

( ) ( )( ) ( )( )[ ]2 2

222

2 2 1, 2

, cos sini i i ii i N N

x yu x v y u x v yu vg x y j eπ π σ

πσ

+−

+ += +




How to go further …?

� Problem: How to segment� textured objects from background� adjacent textures

Each pixel is described by many texture attributes: a feature vectorPixels from a single texture yield a cluster in feature space.

� Solutions:� Unsupervised: Cluster analysis� Supervised: Pattern recognition

allows training a classifier based on examples

texture filtering fourier transform and

Documents