Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo...

19
Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department of Computer Science, University of Joensuu Box 111, Fin-80101 Joensuu FINLAND [email protected].fi

Transcript of Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo...

Page 1: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Random Swap EM algorithm for GMM and Image Segmentation

Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti

Speech & Image Processing UnitDepartment of Computer Science, University of Joensuu

Box 111, Fin-80101 JoensuuFINLAND

[email protected]

Page 2: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Outline

Background & StatusRS-EMApplication

Page 3: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Background: Mixture Model

Page 4: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Background: EM algorithm

EM algorithm -> {α, Θ} E-step (Expectation):

M-step (Maximization):

Iterate E,M step until convergence α- mixing coefficient

Θ- model parameters, eg. {μ,∑}

( 1) ( 1)( , ) [log ( , | ) | , ]i iQ E p X Y X

( ) ( 1)argmax ( , )i iQ

Page 5: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Local MaximaLet’s describe it as mountain climbing……

Page 6: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

600km

2160m 3099m

Page 7: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.
Page 8: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Initialization Effect

Initialization and Result(1) Initialization and Result(2)

Page 9: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Sub-optimal Example

The situation of local maxima trap

Page 10: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Status

Standard EM for Mixture Models(1977) Deterministic Annealing EM (DAEM) (1998) Split-Merge EM (SMEM) (2000) Greedy EM (2002) RS-EM coming…

Page 11: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Outline

Background & StatusRS-EM (Random Swap)Application

Page 12: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

RSEM: Motivations Random manner Prevent from staying near the unstable or

hyperbolic fixed points of EM. Prevent from its stable fixed points corresponding

to insignificant local maxima of the likelihood function

Avoid the slow convergence of EM algorithm Less sensitive to its initialization

Page 13: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Formulas SMEM

Greedy EM

RSEM

Page 14: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Random Swap EMAfter EM

Afte

r Sw

ap

After EM

Page 15: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Comparisons(1)

Page 16: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Comparisons(2)

Q1 Q2 S1 S4

Page 17: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Outline

Background & StatusRS-EMApplication

Page 18: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Application Image Segmentation Color Quantization Image Retrieval ……

Page 19: Random Swap EM algorithm for GMM and Image Segmentation Qinpei Zhao, Ville Hautamäki, Ismo Kärkkäinen, Pasi Fränti Speech & Image Processing Unit Department.

Conclusion Introduce Randomization into algorithm Performs better Without heavy time complexity Wider applications

Thanks!☺