an image registration technique for recovering rotation, scale and translation parameters march 25,...

46
An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

Upload: avery-raynolds

Post on 11-Dec-2015

224 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

An image registration technique for recovering rotation, scale and

translation parameters

March 25, 1998

Morgan McGuire

Page 2: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 2

Acknowledgements

• Dr. Harold Stone, NEC Research Institute

• Bo Tao, Princeton University

• NEC Research Institute

Page 3: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 3

Problem DomainSatellite, Aerial, and Medical sensors produce series images which need to be aligned for analysis. These images may differ by any transformation (possible noninvertible).

Images courtesy of Positive Systems

Page 4: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 4

New Technique

• Solves subproblem (practical case)

• O(ns(NlogN)/4k+Nk) compared to O(NlogN), O(N3)

• Correlations typically > .75 compared to .03

Page 5: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 5

Structure of the Talk• Differences Between Images

• Fourier RST Theorem

• Degradation in the Finite Case

• New Registration Algorithm– Edge Blurring Filter– Rotation & Scale Signatures

• Experimental Results

• Conclusions

Page 6: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 6

Differences Between Images

• Alignment

• Occlusion

• Noise

• Change

Page 7: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 7

Sub-problem Domain

• Alignment = RSTL• Occlusion < 50%• Noise + Change = Small• Square, finite, discrete

images• Image cropped from

arbitrary infinite texture

n

n

N pixels

Page 8: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 8

RST Transformation

1100

cossin

sincos

1p

p

r

r

y

x

yss

xss

y

x

))cossin(,)sincos((, syxysyxxpyxr

Page 9: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 9

Fourier Rotation, Scale, and Translation Theorem†

ssFseF yxyxpsyxj

yxryx /cossin,/sincos, 2/)(

prr FPFRrF ,Let .DTFT Where

Pixel Domain Fourier Domain

p = rotate(r, ) P = rotate(R, )

p = dilate(r, s) Fp = s2 . dilate(Fr, 1/s)

p = translate(r, x, y) Fp = translate(Fr, x, y)

Page 10: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 10

†For Infinite Images

Page 11: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 11

In practice, we use the DFT

Let X0 = DFT(x0)

X0 and x0 are discrete, with N non-zero coefficients.

Let X = DTFT(x)

k

kXX )2(** sin0

k

kNtxx )(*0

X0 and x0 are sub-sampled tiles (one period spans) of X and x. The Fourier RST theorem holds for X and x... does it also hold for X0 and x0?

Page 12: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 12

Fourier Transform and Rotations

Page 13: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 13

Theorem

Infinite case: Fourier transform commutes with rotation

Folklore: It is true for the finite case

Page 14: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 14

Using Fourier-Mellin Theory

• Magnitude of Fourier Transform exhibits rotation, but not translation

• Registration algorithm:– Correlate Fourier Transform magnitudes

for rotation– Remove rotation, find translation

• Generalizes to find scale factors, rotations, and translation as distinct operations

Page 15: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 15

Folklore is wrongImage

Image

Tile

Tile Rotate

Rotate

Page 16: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 16

The Mathematical Proof

F f x y , ,

Transform, then rotate

The Finite Fourier transform

f x y,( , ) F

j x y N2 ( ) / D f x yyx

, ,

Windowing, sampling,

infinite tiling

D f x y eyx

j x y N , , ( )/2

continuous

Page 17: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 17

The Mathematical ProofRotate, then transform

F f x y , ,

D f x y eyx

, , j x y N2 /

D f x y eyx

, , j x y N2 /

x y

D f x y e j x y N , ( , ) ( )/2

Page 18: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 18

Finite-Transform Pairs

Page 19: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 19

The Artifacts

Page 20: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 20

Fourier Transforms

ts.coefficien nonzero with finite

are they so,~

,~ of periods single are and :

sampled.-sub is ~

. period ~

,~

][~][~

;][~

][~:

][;][:

;:

00

1

0

/21

0

1/21

21

21

N

XxXxDFT

XXNXx

enxkXekXnxDFS

enxXdeXnxDTFT

dtetxXdeXtxCTFT

N

n

NnkjN

kN

Nknj

N

n

njnj

jtjt

Oppenheim & Willsky Signals & Systems; Oppenheim and Schafer, Discrete-Time Signal Processing

Page 21: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 21

Tiling does not Commute with Rotation

Tiled Image Rotated Tiled Image Tiled Rotated Image

…so the Fourier RST Theorem does not hold for DFT transforms.

Page 22: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 22

Correlation Computation

( , )x y

x yN

x y

xN

x yN

y

i i i i

i i i i

1

1 122

22

xy

x y

Page 23: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 23

Prior Art

• Alliney & Morandi (1986) – use projections to register translation-only in

O(n), show aliasing in Fourier T theorem

• Reddy & Chatterji (1996) – use Fourier RST theorem to register in

O(NlogN)

• Stone, Tao & McGuire (1997) – show aliasing in Fourier RST theorem

Page 24: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 24

An Empirical ObservationEven though the Fourier RST Theorem does not hold for finite images, we observe the DFT does have a “signature” that transforms in a method predicted by the Theorem.

Image

DFT Magnitude

Page 25: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 25

Sources of Degradation

• Frequency– Aliasing (from Tiling)

– “+” Artifact

– Sampling Error

• Pixel– Image Window Occlusion

– Image Noise

Page 26: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 26

Algorithm Overview

Norm. Circ. Corr.

r p

G G

FMT

f,d f,d

f,logd

f,logd

J

Maximum Value Detector

Peak Detector

Norm. Corr.

List of scale factors (s)

exp

J

FMT

W W

HH

Coarse (x, y)

FFTFFT

Dilate

Rotate

FFT

Dilate

Rotate

FFT

(Pixel) Correlation

W W W W

r m p h

1. Pre-Process

3. Recover

Scale

Parameter

4. Recover

Rotation

Parameter

5. Recover Translation Parameters

2. FMLP Transform

Page 27: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 27

Problem: “+” Artifact

None Rotation Dilation TranslationTransformation

DFT

Image

Page 28: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 28

Solution: “Edge-Blurring” Filter, G

Image

None

DFT

Disk BlurFilter

Page 29: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 29

Problem:Need Orthogonal Invariants

dxdyeyxrG

RR

yxj

yx

)(,

,sin,cos

In the “log-polar” (log,) domain:

Added NoiseTranslate

AliasingNoise,,logby Translateby Dilate

AliasingNoise,,by Shift Cyclicby Rotate

ss

Domain FMLPDomain Pixel

Fourier-Mellin transform:

Page 30: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 30

Mapping (x,y) to (log,)

y

x

x=8

y=8

log

log

/4

log

/4

x=4

y=4

Page 31: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 31

Sample Image Pair

G(r) G(p)

= 17.0o

s = 0.80

x = 10.0

y = -15.0

N = 65536

k = 2

Page 32: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 32

Nonzero Fourier Coefficients

R P

Page 33: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 33

Solution I: Rotation Signature

2/

0

sin,cosn

r dRJ

1. Selectively weight “edge coefficients” (J filter)

2. Integrate along axis

is Scale and Translation Invariant. Pixel rotation appears as a cyclic shift => use simple 1d O(nlogn) correlation to recover rotation parameter.

Page 34: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 34

Signatures of r and p

Page 35: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 35

Correlations

Page 36: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 36

Solution II: Scale Signature

0

1 sin,coslog dRHSr

1. Integrate along axis (rings)

2. Normalize by (area)

3. Enhance S/N ratio (H filter)

S is Rotation and Translation Invariant. Pixel dilation appears as a translation => use simple 1d O(nlogn) correlation to recover scale parameter.

Page 37: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 37

Raw S Signature

Page 38: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 38

Filtered S Signature

Page 39: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 39

S Correlation

Page 40: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 40

New Registration Algorithm

Norm. Circ. Corr.

r p

G G

FMT

f,d f,d

f,logd

f,logd

J

Maximum Value Detector

Peak Detector

Norm. Corr.

List of scale factors (s)

exp

J

FMT

W W

HH

Coarse (x, y)

FFTFFT

Dilate

Rotate

FFT

Dilate

Rotate

FFT

(Pixel) Correlation

W W W W

r m p h

Compute full-resolution Correlation for small neighborhood of Coarse (x, y) to refine.

Page 41: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 41

Recovered Parameters s (x, y)

Fish Actual 17.00 0.80 (10.00, -15.00)(natural) Recovered 17.25 0.75 (10.17, -14.98)

Peak Corr. 0.82 0.53 0.90

1d RMS = 3.42 2d RMS = 4.84

Image

Page 42: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 42

Disparity Map

Page 43: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 43

Multiresolution for Speed

• Algorithm is O(NlogN) because of FFT’s

• With kth order wavelet, O((NlogN)/4k)

• To refine, search 22k = 4k positions

• Using binary search, k extra trials @ O(N) each

• Total algorithm is O((NlogN)/4k + Nk)

Page 44: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 44

Results & ConfidenceTrial s (x, y)

1 suburban Actual 31.00 1.00 (0.00, 0.00)(aerial) Recovered 31.00 1.00 (0.00, 0.00)

Peak Corr. 0.90 0.48 0.972 suburban Actual 0.00 1.40 (0.00, 0.00)

Recovered 0.25 1.37 (0.00, 0.00)Peak Corr. 0.86 0.31 0.84

3 suburban Actual 0.00 1.00 (10.00, -3.00)Recovered 0.00 1.00 (10.00, -3.00)Peak Corr. 0.91 0.83 1.00

4 fish Actual 24.00 0.70 (7.00, 12.00)(natural) Recovered 25.00 0.72 (7.06, 11.99)

Peak Corr. 0.78 0.45 0.955 mamogram Actual -10.00 1.20 (10.00, 9.00)

(x-ray) Recovered -10.25 1.15 (9.76, 8.75)Peak Corr. 0.94 0.73 0.58

6 essai Actual -15.00 1.50 (-10.00, -10.00)(satellite) Recovered -15.25 1.56 (-11.33, -11.40)

Peak Corr. 0.94 0.40 0.78

Image

Page 45: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 45

Analysis of Results

s (x, y) TotalMean Correlation 0.89 0.53 0.85 0.76

Mean RMS Error (1d) 1.92

Page 46: An image registration technique for recovering rotation, scale and translation parameters March 25, 1998 Morgan McGuire

3/25/98 Morgan McGuire 46

Future Directions

• Better scale signature

• Use occlusion masks for FM techniques?

• Combining FM technique with feature based techniques