U E D R , A S I H C RSS

논문번역/2012년스터디/서민관

off-line

abstract
off-line .
(segmentation-free approach) .
, , 통 , , , 한 .
히, 형 판(linear discriminant analysis) allograph , 통 합한 .
(lexicon-off) .

1.
.
off-line .
, .
.
.

off-line 한 Hidden-Markov-Model .
, .
, 형 판, allograph , 통 .

.
3 .
, , 통 .
한 평 7 .

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

.
. .
.
(feature extract) .
1, 18 15 HMMs recurrent neural network HMMs 합한 .
15 행해 , 1, 18 .
off-line 16 .

, 9 할(segmentation-free) .
, 통 .
한 향 10 11 .
, .
, HMMs allograph 형 판 한 효 .

3.
Lancaster-Oslo/Bergen .
IAM Bern 형태 .
한 텍(, , , ) 500 한 1200 형태 .
해 250 a..f 형태 . 해 6 c03 .

(Senior) 형태 .
한 25페 , .

형태 300dpi 256 .
Fig. 1 .

4.
형태 확한 .
.
IAM . 통해 .

.
할하 .
.
Otsu method .
평 투 . .

.
, .
.

, .
통해 .
.

15 형 회(linear regression) . .
, -흰, 흰- .
.
(slant angle) .

화하 (local extrema) .
(scaling factor) . .
.
0 255 .
3 .

5.
.
11 sliding window .
4 2 . 하학 .

sliding window 7 .
(1) -흰 (window 화 한 )
(2) (position of the mean value of the intensity distribution)
(3) 텍트 픽
(4) 텍트 픽
(5) 텍트 픽
(6) 텍트 픽
(7)
(2)-(5) ( 한 line fitting ) .
window 4 .

, 3 .
window 4 , 평 . (8), (9), (10) .

, 평 파(approximate horizental derivative) . 20 .(window 10 + 10)

(...) 형 판 통합하.(cf. 6)
(........ original feature representation) 화하.
환 A class scatter matrix Sw scatter matrix Sb (eigenvalue) .
scatter matrix HMM 행해. (...........................)
scatter matirx LDA .

...

... ... S-1wSb .
m 한 m .
LDA 환 한 HMM 학 .

6. 통
, 학, HMMs ESMERALDA 5 .
HMMs 512 Gaussian mixtures with diagonal covariance matrices codebook - .
52 , 10 , 12 호, 하 Baum-Welch .
.
Viterbi beam-search .

13 allograph .
Allograph 플(realization) .
HMMs .
allograph HMMs (heuristically) . allograph .
, 훈 allograph HMMs .
, allograph .
.
allograph soft vector .

, 통합.
... x 화 하 W^ .
P(w) w x 할 확 P(x|w) .
absolute discounting 한 bi-gram backing-off for smoothing of probability distributions .

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

해 IAM c03 440 , 109 .
6 .
LDA( 12) 한 14.2% . allograph ( 6 allograph) 13.3% .
bi-gram 11.1%(테 perplexity 12.0) .
39.0%, 421 (호 포함) 포함하 13.9% . 11 20.5% .

행하 .
IAM a-f( 250 ) , 4321트(a-d ) , 1097트(e-f ) .
31.3% .
allograph 한 향 .
3 allograph 31.1% , 10 allograph (34.8%) .
LDA 29.1% .
포함 22.2% .(테 perplexity 12.0)
60.6%.

8. .
off-line . , , 행하.
.
히 향. 형 판 행하 .
allograph , 히, .

9. acknowledgement
German Research Foundation(DFG) 트 Fi799/1 .
형태 DB .
Valid XHTML 1.0! Valid CSS! powered by MoniWiki
last modified 2021-02-07 05:28:54
Processing time 0.1086 sec