Handwritten character recognition Handwritten character recognition - PowerPoint PPT Presentation

Handwritten character recognition Handwritten character recognition using elastic matching based on using elastic matching based on a class- -dependent deformation model dependent deformation model a class S. Uchida and H. Sakoe Kyushu University, Japan

Introduction Introduction

Elastic matching Elastic matching displacement field v input pattern input fitted to reference reference 3 Human Interface Lab. Kyushu Univ.

Conventional EM techniques Conventional EM techniques = category-independent = it is assumed that all classes have the same deformation tendency however frequent rare may cause overfitting to “A” ! 4 Human Interface Lab. Kyushu Univ.

Our purpose Our purpose develop new EM where class-dependent deformation tendency(= eigen-deformation) is considered 5 Human Interface Lab. Kyushu Univ.

Two tasks Two tasks � How to get the eigen-deformations? → Reported already [Uchida and Sakoe, Pattern Recognition , 2003] � How to utilize the eigen-deformations in EM? 6 Human Interface Lab. Kyushu Univ.

Estimation of Eigen-Deformations eigen-deformations collection of displacement fields using (conventional) EM … Principal Component Analysis … … reference training patterns 7 Human Interface Lab. Kyushu Univ.

Estimated eigen- -deformations deformations Estimated eigen u 1 u ２ u ３ apply 0 apply negatively positively 8 Human Interface Lab. Kyushu Univ.

Two tasks Two tasks � How to estimate the eigen-deformations? → Reported already [Uchida and Sakoe, Pattern Recognition , 2003] � How to utilize the eigen-deformations in EM? → Topic of this presentation 9 Human Interface Lab. Kyushu Univ.

Outline of Outline of proposed technique proposed technique

Class- -dependent deformation model dependent deformation model Class Deformation model of reference pattern Pc ( x,y ) ( class c ):   M + ∑ ( ) ( )   α P x y u x y , , c m c m ,   = m 1 = α 1 × + α 2 × …+ α M × u c,1 u c,2 u c,M 11 Human Interface Lab. Kyushu Univ.

Geometric interpretation Geometric interpretation   M + ∑ ( ) ( )   α P x y u x y , , c m c m ,   = m 1 M-dimensional manifold ... > α 0 m < α P 0 m c = α ( 0 ) m pattern space 12 Human Interface Lab. Kyushu Univ.

EM based on class- -depend. deform. model depend. deform. model EM based on class 2     M ∑ ( ) ( ) ( ) ∫   − + α   E x y P x y u x y dx dy , , ,   c m c m ,     = m 1 → α minimize w.r.t. m α nonlinear m ... function reference of α P c input E pattern space 13 Human Interface Lab. Kyushu Univ.

Solution via linear approximation Solution via linear approximation nonlinear Taylor expansion ... & approx. P reference c linear !! Tangent plane input ( M -dimensional) E Easily solvable as a least square problem 14 Human Interface Lab. Kyushu Univ.

Experimental Results Experimental Results

Database Database English capitals from ETL6 1100 samples × 26classes preprocessing linear scaling, etc. # 1-100 # 101-600 # 601-1100 average reference training patterns test patterns to estimate eigen-d in recog. exp. 16 Human Interface Lab. Kyushu Univ.

Recognition rate Recognition rate 3 eigen-deforms 99.3% were enough! 99.2% 99.1% 99.0% 98.9% 98.8% 0 10 20 30 40 50 # eigen-deformations, M 17 Human Interface Lab. Kyushu Univ.

Comparison to class- -in independent EM (1) dependent EM (1) Comparison to class 0.23% 99.3% （ ３ / ４ misrecognitions ） recognition rate 99.2% 99.1% affine EM 99.0% (M=6) 98.9% 98.8% 0 10 20 30 40 50 # eigen-deformations, M 18 Human Interface Lab. Kyushu Univ.

Conclusion & Future work Conclusion & Future work � Conclusion � EM based on a class-dependent deformation model was developed � High accuracy and efficiency were shown through character recognition experiment � Future work � Solution strategies other than linear approx. � Relation to the sub-space methods 19 Human Interface Lab. Kyushu Univ.