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

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) displacement deformation-invariant field distance reference input warped reference 2 Human Interface Lab. Kyushu Univ.

Overfitting Problem Overfitting Problem reference input input “R” may be misrecognized as “A” 3 Human Interface Lab. Kyushu Univ.

Purpose Purpose Reduce the misrecognition due to overfitting by using eigen-deformations not eigen-deform. eigen-deform. 4 Human Interface Lab. Kyushu Univ.

Two Central Problems Two Central Problems � How to estimate eigen-deformations ? � How to use the eigen-deformations in EM-based recognizer ? 5 Human Interface Lab. Kyushu Univ.

Estimation of Estimation of Eigen- -Deformations Deformations Eigen

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

Piecewise- -Linear EM Linear EM Piecewise reference training pattern displacement field linear interpolation constraints for topology preservation 8 Human Interface Lab. Kyushu Univ.

st Eigen Estimated 1 st Eigen- -Deformations Deformations Estimated 1 「 Ａ 」 「 Ｂ 」 「 Ｃ 」 「 Ｄ 」 0 apply apply (reference) positively negatively 9 Human Interface Lab. Kyushu Univ.

Cumulative Proportion Cumulative Proportion Top % 30 100 20 80 10 60 5 3 40 20 1 0 over 50% with 3-5 (of 74) eigen-deformations 10 Human Interface Lab. Kyushu Univ.

Recognition Using Eigen- - Recognition Using Eigen Deformations Deformations

Recognition Using Eigen- -Deformations Deformations Recognition Using Eigen input reference image eigen-deform. of “A” disp. field v distance … + Mahalanobis min dist. distance discrimination 12 Human Interface Lab. Kyushu Univ.

Recognition Result (1) Recognition Result (1) with eigen-deform. (proposed) 9 9 98.94 recognition rate (%) 500 samples/category 9 8 . 5 42% reduction of misrecognitions 9 8 98.18 9 7 . 5 without 97.12 9 7 eigen-deform. (conventional) 9 6 . 5 0 1 2 3 4 5 maximum displacement (pixels) 13 Human Interface Lab. Kyushu Univ.

Effect on Overfitting Reduction Effect on Overfitting Reduction with eigen-deform. (proposed) 9 9 98.94 recognition rate (%) 9 8 . 5 degradation by the increase of overfitting 9 8 98.18 9 7 . 5 without 97.12 9 7 eigen-deform. (conventional) 9 6 . 5 0 1 2 3 4 5 maximum displacement (pixels) 14 Human Interface Lab. Kyushu Univ.

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 15 Human Interface Lab. Kyushu Univ.

Comparison with Another Evaluation Method Comparison with Another Evaluation Method with eigen-deform. (amplitude + direction) 9 9 98.94 recognition rate (%) 9 8 . 5 9 8 98.18 with amplitude 9 7 . 5 of displacement field 97.12 9 7 9 6 . 5 0 1 2 3 4 5 maximum displacement (pixels) 16 Human Interface Lab. Kyushu Univ.

Data Data English capital letters from ETL6 (1100 samples / category) preprocessing (size normalization, blurring, histogram equalization…) #1-100 #101-600 #601-1100 average reference training patterns test patterns 17 Human Interface Lab. Kyushu Univ.

Mahalanobis Distance Mahalanobis Distance M Σ 1 〈 v - v , u 〉 λ c c,m c,m m =1 v : displacement field to be evaluated c : class (“A”, “B”, ..) of reference u c,m : m -th eigen-deformation of class c λ c,m : contribution (eigenvalue) of u c,m 18 Human Interface Lab. Kyushu Univ.

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 pivots on input warp on warped A A B B B given results 19 Human Interface Lab. Kyushu Univ.

Effect on Overfitting Reduction (2) Effect on Overfitting Reduction (2) the most remarkable improvement: M � H (30 misrecognitions � 13) in [Ronee et al., ICDAR2001] “M � H is the most typical misrecognition due to overfitting.” 20 Human Interface Lab. Kyushu Univ.