Bradford University & Meudon

naivenorthΤεχνίτη Νοημοσύνη και Ρομποτική

8 Νοε 2013 (πριν από 3 χρόνια και 11 μήνες)

64 εμφανίσεις

Survey of feature recognition
techniques


Work package 5


Bradford University & Meudon
Observatory


V V Zharkova, S S Ipson



SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

Summary of the recognition techniques


Method

Feature

Histogram
method

LOG

Region
Growing

Simulated
Annealing

Baysian
Inference

ANN

Hough
Transform

Valley
Detection

MALM

Sunspots

P/I

P/I

P/A

P/I

Filaments

P/A

I

P/I

Plage

P/I

A, P/A

P/A

CMEs

P/I

P/A

P/I

Emerging

MagnFlux

P/A

Coronal

Holes

Waves

Flares

Bright

Points

P
-

pre
-
processing
,

I
-

user interaction was required

and

A

-

automated method

SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


2.1 Histogram
-
based segmentation



2.2. Region
-
based segmentation


2.2.1 Region growing


2.2.2 Clustering


2.2.3 Multi
-
resolution transforms


2.3 Edge
-
based segmentation


2.3.1 Gradient operator based edge detection


2.3.2 Canny edge detection


2.3.3 Laplacian of Gaussian zero
-
crossing edge
detection


2.4 Artificial neural networks


2.4.1 Standard technique



2.4.2 Cascade
-
correlation architecture


2.4.3 Evolving cascade neural networks


2.4.4 GMDH
-
type neural networks


2.4.5 Generalized regression neural networks


2.5 Explicit
-
model based segmentation


2.5.1 The Hough transform


2.5.2 Ribbon detection


2.6 Models based on functionals


2.6.1 Active contours


2.7 Bayesian inference


2.8 Motion segmentation



2.9 Shape analysis


2.10 Classification


II. Survey of Pattern Recognition Techniques

3.1
Image preparation

3.1.1 Geometrical distortion

3.1.2 Blurring


3.1.3 Intensity calibration

3.1.4 Miscellaneous defects

3.2 Detection of sunspots


3.2.1 Histogram methods


3.2.2 LOG methods


3.2.3 Region growing methods


3.2.4 Simulated annealing

3.3 Filament detection


3.3.1 Chain linking procedure


3.3.2 Region growing procedure

3.4 Detection of active regions (plage)


3.4.1 Global intensity threshold


3.4.2 Region growing methods


3.4.3 Bayesian inference method

3.5 Detection of coronal mass ejections


3.5.1 Hough transform method


3.5.2 Multiple abstraction level mining



method

SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

III.
a. Why the pre
-
processing techniques?


Difficulties with images:


Errors

in FITS header information


Image
shape

(ellipse),
centre

and
the
pole
coordinates


Weather
transparency

(clouds)
and different
thickness
of
atmosphere


Centre
-
to
-
limb
darkening


Defects

in data (strips, lines,
intensity)


SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

SUNSPOTS

Synoptic Charts

Central Meridian Synoptic Map

SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

Image segmentation procedures







Thresholding approaches (histogram
-
based segmentation)



Edge
-
based methods (using the first or second derivatives of the spatio
-
temporal functions



Region growing methods (intitial starting pixel + criterion for merging)



Hybrid region growing and edge detection techniques



Neural networks (training without explicit criteria)



Global Information methods (Bayesian, functional models, Hough
transform)



Miscellaneous (data clustering, simulated annealing, data mining)

SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


General techniques



Histogram
-
based segmentation




Analyse the grey
-
level histograms



Size of the segmented object varies with the threshold


Give good results on a uniform background


Objects had a distinct intensity range



Region
-
based segmentation


Region growing (start from seeds and grow regions on specified criteria)



Clustering (pixels are clustered in a feature space using any discriminating
feature asociated and then connecting regions are found)



Edge
-
based segmentation



Relies on discontinuities in the image data to locate boundaries


But edge profile is not known


Profile can vary with edge (shading or texture)



SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


Edge
-
based segmentation



Gradient operator based edge detection




Vertical and horizontal components are finite difference formulae with


Sobel convolution masks: vertical and horizontal






-
1
-
2
-
1
-
1 0 1







0 0 0
-
2 0 2







1 2 1
-
1 0 1


Gradient magnitude
-

a square root of the sum of the square gradient components


Candidate edge located with gradient magnitude above threshold


Multi passes of the detected edge


Canny edge detection


Smooth image with a Gaussian filter



Compute gradient magnitude and orientation with finite differences


Apply non
-
maxima suppression to thin the gradient
-
magnitude edge image


Track along edges starting from the point esceeding higher threshold with the edge point
esceeding the lower threshold


Apply edge linking to fill small gaps












SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


Edge
-
based segmentation




Laplacian of Gaussian zero
-
crossing edge detection (LOG)



The Laplacian
-

2D isotropic measure of the second spatial derivative of an image


L of an image has the lagest magnitudes at peaks of intensity


L of an image has zero crossings at the points of inflection


Common convolution kernels to calculate digital Laplacian:







0 1 0 1 1 1







1
-
4 1 1
-
8 1







0 1 0 1 1 1



L sensitive to noise => applied after a Gaussian smoothing filter


Hence => LOG or Marr
-
Hildreth operator









SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


Explicit model
-
based segmentation




The Hough transform (CMEs


Bergmans)



Uses an accumulator array with dimension equal the number of parameters in the
family of curves to be detected


If
y
=
ax + b, then a and b and accumulator array indices (2) correspond


Accumulator array



Ribbon detection



Modified Hough transform which includes a directions of the intensity gradient


across the line or curve









SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

Miscelleneous methods


Image cleaning (solar: shape and intensity)


Image filtering


Image enhancement (to increase a contrast)


Morphological operations (to complete the
feature shape)


Others (reported by other speakers)



SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


Artificial Neural Networks




Standard technique



Exploits a feed
-
forward fully connected network: input, hidden or output neurons


connected by adjustable synaptic weights


The technique implies that ANN structure is well defined


It means that one must preset the input and hidden neurons


Apply suitable neuron activation function


Sigmoid activation function:





y
=
f
(
x, w
) = 1/(1 + exp(


w0


Σ
i
m

wi x
i)),







where m


number of variables x
1
, xm, X is the input vector,
w
is a synaptic weigh
vector


User must choose a suitable learning algorithm


Rationally set learning rate, a number of the training epochs etc.


If ANN includes 2 hidden neurons
-
> back
-
projection algorithm provides best results










SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

Filament recognition with ANN




f
1

f
2

f
3

z
1
(
j
)

z
2
(
j
)

z
r
(
j
)

(
j
)

w
0
(2)

w
0
(1)

w
0
(3)

u
j

s
j

y
j

= (0, 1)





1(b)
0
5000
10000
0
500
1000
1500
2000
Y
0
5000
10000
-4
-2
0
2
4
F
1(a)
SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

Recognised filaments

1(a)
1(b)
SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

Method

Feature

Histogram
metho
d

LOG

Region
Gro
wing

Simulate
d
Ann
ealin
g

Baysian
Infer
ence

ANN

Hough
Tr
an
sfo
rm

Valley
Dete
ction

MALM

Sunspots

P/I

P/I

P/A

P/I

Filaments

P/A

I

P/I

Plage

P/I

A, P/A

P/A

CMEs

P/I

P/A

P/I

Magnetic

field


P/A



P/I

Summary of the Solar Feature Recognition Methods

SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03

VII. Conclusions


WP5 is successfully implementing the project plan



Feature recognition in solar images generated a substantial
interest among the IT and solar community
-
FR Workshop



A few novel techniques were developed for each feature
(see sunspots, ARs, filaments (ANN + MO), magnetic NL)



Ongoing collaboration with the partners from Meudon,
NSO, UAS, IAS and OATO



The current status


a detailed catalogue design stage

SFR Workshop 1, BRO, Brussels,
23
-
24 Oct '03


WP5

Feature Recognition

Work in progress




Adjustment of the FR techniques to the specifics of each
catalogue with respect to the time coverage period and
providers for the Unified Observing Catalogues (UOC)




Created an Access database fed by the detected sunspot
feature parameters and developed a preliminary query and
response pages



Preparing a Demo on the Web for your testing



http://www.cyber.brad.ac.uk/egso/