U E D R , A S I H C RSS

논문번역/2012년스터디/서민관

off-line

abstract
off-line .
(segmentation-free approach) .
, , 고, , , .
, (linear discriminant analysis)과 allograph , .
(lexicon-off) 결과 .

1.
.
off-line .
고, .
경계 .
.

off-line Hidden-Markov-Model .
, .
, , allograph , .

.
3 .
, , .
결과 7 .

2.
off-line .
legal amount reading 고, . 그 2, 8.

.
구간 . .
.
(feature extract) .
1, 1815 HMMs 과 recurrent neural network HMMs .
15 , 1, 18 .
off-line 16 .

, 9 (segmentation-free) .
고, 결과 .
10 11 .
, .
, 각 HMMs allograph .

3.
Lancaster-Oslo/Bergen .
IAM과 Bern .
(, 교, , )과 500 1200 .
250 a..f . 그 6 c03 .

(Senior) .
25 , 공개 .

300dpi 256 .
Fig. 1 .

4.
.
.
IAM . .

각각 .
.
.
Otsu method .
. 그 .

.
, .
.

, .
.
.

15 (linear regression) . .
고, -, - .
.
균값 (slant angle) .

극값(local extrema) .
(scaling factor) . .
.
간격 0 255 .
3 .

5.
.
11 sliding window 기 .
4 2 . 그 .

sliding window 7 .
(1) - (window )
(2) 기 극값 균값 (position of the mean value of the intensity distribution)
(3)
(4)
(5)
(6) 균 극값
(7) 균 극값
견고 (2)-(5) ( 고값 line fitting) .
window 4 .

극값 균값 , 3개 .
window 4개 , 균 값 . 그 각각 (8), (9), (10) .

, (approximate horizental derivative). 20 .(window 10개 + 10개 )

(...) .(cf. 6)
공간 (........ original feature representation) .
A class scatter matrix Sw과 scatter matrix Sb 간값(eigenvalue) .
scatter matrix HMM . (...........................)
scatter matirx LDA .

...

... ... S-1wSb값과 고.
m개 m개 .
LDA HMM .

6.
, , HMMs ESMERALDA 개 5 .
HMMs 512개 Gaussian mixtures with diagonal covariance matrices codebook과 - .
52개 , 10개 , 12개 , Baum-Welch .
.
Viterbi beam-search .

13 allograph .
Allograph (realization)과 .
HMMs .
allograph HMMs (heuristically) 결. allograph .
, allograph HMMs .
, allograph .
.
allograph 게 결 soft vector .

, .
... x W^ .
P(w) w x P(x|w) .
absolute discounting bi-gram 과 backing-off for smoothing of probability distributions .

7. 결과
. , , .
table 1 . 고, , bi-gram 결과 .
IAM a..d 고, .
(lexicon-free ....) 결과 table 2 .
(Senior) 282 141 . bi-gram perplexity 15.3.
bi-gram 13.3% 12.1%까 결과 .
LDA . LDA 공간 12까 .
(table 2) 28.5%, 1.5k 10.5%.
결과 결과(literature ......) . .
17 28.3% . 84.1%고, 1.3k 16.5%.
15 6.6% 41.1% .
9 15.0%.

IAM c03 440 고, 109 .
6 .
LDA( 12) 14.2% . allograph (각 6개 allograph) 13.3%까.
bi-gram 11.1%( perplexity 12.0)까 .
39.0%, 421개 ( ) 13.9%까 . 11 20.5% .

결과 .
IAM a-f( 250 ) , 4321 (a-d ) 고, 1097 (e-f ) .
31.3% .
allograph .
3개 allograph 31.1% 고, 10개 allograph (34.8%)과 .
LDA 29.1% 게 감.
22.2% .( perplexity 12.0)
60.6%.

8. 결.
off-line . 그 , , .
결과 .
. 그 공간 .
allograph , , .

9. acknowledgement
German Research Foundation(DFG) Fi799/1 .
DB 게 감 .
Valid XHTML 1.0! Valid CSS! powered by MoniWiki
last modified 2021-02-07 05:28:54
Processing time 0.0563 sec