Using Eigen- -Deformations in Deformations in Using Eigen - - PowerPoint PPT Presentation

▶
using eigen deformations in deformations in using eigen
SMART_READER_LITE
LIVE PREVIEW

Using Eigen- -Deformations in Deformations in Using Eigen - - PowerPoint PPT Presentation

Mar 06, 2023 259 likes •467 views

Using Eigen- -Deformations in Deformations in Using Eigen Handwritten Character Recognition Handwritten Character Recognition S. Uchida M. A. Ronee H. Sakoe Kyushu University Fukuoka, Japan Elastic Matching (EM) Elastic Matching (EM)

slide-1
SLIDE 1

Using Eigen Using Eigen-

  • Deformations in

Deformations in Handwritten Character Recognition Handwritten Character Recognition

  • S. Uchida
  • M. A. Ronee
  • H. Sakoe

Kyushu University Fukuoka, Japan

slide-2
SLIDE 2

Human Interface Lab. Kyushu Univ.

2

Elastic Matching (EM) Elastic Matching (EM)

input reference displacement field deformation-invariant distance warped reference

slide-3
SLIDE 3

Human Interface Lab. Kyushu Univ.

3

Overfitting Problem Overfitting Problem

reference input

input “R” may be misrecognized as “A”

slide-4
SLIDE 4

Human Interface Lab. Kyushu Univ.

4

Purpose Purpose

Reduce the misrecognition due to overfitting by using eigen-deformations eigen-deform. not eigen-deform.

slide-5
SLIDE 5

Human Interface Lab. Kyushu Univ.

5

Two Central Problems Two Central Problems

How to estimate eigen-deformations ? How to use the eigen-deformations in EM-based recognizer ?

slide-6
SLIDE 6

Estimation of Estimation of Eigen Eigen-

  • Deformations

Deformations

slide-7
SLIDE 7

Human Interface Lab. Kyushu Univ.

7

Estimation of Eigen Estimation of Eigen-

  • Deformations

Deformations

training patterns reference collection of displacement fields using EM

…

Principal Component Analysis

… …

eigen-deformations

slide-8
SLIDE 8

Human Interface Lab. Kyushu Univ.

8

Piecewise Piecewise-

  • Linear EM

Linear EM

reference training pattern displacement field linear interpolation constraints for topology preservation

slide-9
SLIDE 9

Human Interface Lab. Kyushu Univ.

9

Estimated 1 Estimated 1st

st Eigen

Eigen-

  • Deformations

Deformations

「 Ａ 」 「 Ｂ 」 「 Ｃ 」 「 Ｄ 」

(reference) apply positively apply negatively

slide-10
SLIDE 10

Human Interface Lab. Kyushu Univ.

10

Cumulative Proportion Cumulative Proportion

1 3 5 10 20 Top 30 % 100 80 60 40 20

  • ver 50% with 3-5 (of 74) eigen-deformations
slide-11
SLIDE 11

Recognition Using Eigen Recognition Using Eigen-

  • Deformations

Deformations

slide-12
SLIDE 12

Human Interface Lab. Kyushu Univ.

12

Recognition Using Eigen Recognition Using Eigen-

  • Deformations

Deformations

image distance +

min dist. discrimination reference input

…

  • disp. field v

eigen-deform. of “A” Mahalanobis distance

slide-13
SLIDE 13

Human Interface Lab. Kyushu Univ.

13

Recognition Result (1) Recognition Result (1)

97.12 98.18 98.94

9 6 . 5 9 7 9 7 . 5 9 8 9 8 . 5 9 9

1 2 3 4 5

maximum displacement (pixels)

recognition rate (%)

500 samples/category

without eigen-deform. (conventional) with eigen-deform. (proposed)

42% reduction of misrecognitions

slide-14
SLIDE 14

Human Interface Lab. Kyushu Univ.

14

97.12 98.18 98.94

9 6 . 5 9 7 9 7 . 5 9 8 9 8 . 5 9 9

1 2 3 4 5

maximum displacement (pixels) recognition rate (%)

without eigen-deform. (conventional) with eigen-deform. (proposed)

Effect on Overfitting Reduction Effect on Overfitting Reduction

degradation by the increase of overfitting

slide-15
SLIDE 15

Human Interface Lab. Kyushu Univ.

15

Conclusion and Future Work Conclusion and Future Work

Conclusion

Proposition of the use of eigen-deformations

in EM-based recognizer

Verification of its usefulness through

experiments

Future work

Use of other EM techniques Direct incorporation of eigen-deformations

into EM

slide-16
SLIDE 16

Human Interface Lab. Kyushu Univ.

16

Comparison with Another Evaluation Method Comparison with Another Evaluation Method

97.12 98.18 98.94

9 6 . 5 9 7 9 7 . 5 9 8 9 8 . 5 9 9

1 2 3 4 5

maximum displacement (pixels) recognition rate (%)

with amplitude

  • f displacement field

with eigen-deform. (amplitude + direction)

slide-17
SLIDE 17

Human Interface Lab. Kyushu Univ.

17

Data Data

English capital letters from ETL6 (1100 samples / category) preprocessing

(size normalization, blurring, histogram equalization…) #101-600 average #601-1100 #1-100

reference training patterns test patterns

slide-18
SLIDE 18

Human Interface Lab. Kyushu Univ.

18

Mahalanobis Distance Mahalanobis Distance

M

Σ

m=1

λ

c,m

1

〈 v- v , u 〉

c,m c

v : displacement field to be evaluated

c : class (“A”, “B”, ..) of reference

uc,m: m-th eigen-deformation of class c

λc,m: contribution (eigenvalue) of uc,m

slide-19
SLIDE 19

Human Interface Lab. Kyushu Univ.

19

P i e c e w i s e l i n e a r e l a s t i c m a t c h i n g P i e c e w i s e l i n e a r e l a s t i c m a t c h i n g

reference A pivots on A input B

given

warp on B warped B

results

slide-20
SLIDE 20

Human Interface Lab. Kyushu Univ.

20

Effect on Overfitting Reduction (2) Effect on Overfitting Reduction (2)

the most remarkable improvement:

MH

(30 misrecognitions 13)

in [Ronee et al., ICDAR2001]

“MH is the most typical misrecognition due to overfitting.”