U E D R , A S I H C RSS

논문번역/2012년스터디/이민석


Warning:
ERROR: LaTeX does not work properly.
in /var/www/html/plugin/processor/latex.php on line 210

  • LaTeX

see also
coarse ,

2012/11/16

  • Experiments in Unconstrained Offline Handwritten Text Recognition
  • 1학 1학 Linear Algebra and Its Applications 1.10, 2.1, 2.2

2012/11/29



Experiments in Unconstrained Offline Handwritten Text Recognition( 험)



. . , (feature extraction), 통 형화 , , 험하. 형 판 (Linear Discriminant Analysis), (allograph) 형, 통 통합 .

1
. . 14 (unconstrained) . . , .
한, 휘(lexicon)-free , 행했. 형 판 , , 통 . 토한. 3 . , , 통 . 평 해 7 .

2
. legal amount 한, . 2 8
. . . 1, 9, 11, 18, 15, 17 . 형(HMM) 환형 HMM 1, 18 15 . 15 행한 1, 18 . 16 .
할하 9 , . 합하 . 11 , . . 형, HMM 형 판 (LDA) .

3 (corpus)
Lancaster-Oslo/Bergen . 7 Bern IAM 형태 10 행하. (, , 화, 픽...) 포함하 500 1200 형태 . 250 a..f form 한 하합 c03 .
해 Senior 15 행했. 25 .1 256 300dpi 해 . 1 .
1 ftp://svr-ftp.eng.cam.ac.uk/pub/data

4
확한 "(drift)" . . 4 IAM 확히 .
. (core region) . 핵 , threshold . threshold foreground 픽 . threshold 화한 Otsu 12 . 평 투 minima .
화하 화한. , , (slant) . .
, , . segment . 해 threshold .
15 형 회(linear regression) , edge . 평 흑- -흑 stroke . canny edge detector edge orientation . 히 .
화하 (local extrema) . (scaling) .
화하 한 펜 . 0 255 . 3 .

5
. sliding window 11 . .
sliding window 7 . (1) 흑- (windowed text image ), (2) , (3) , (4) , (5) 트 픽 , (6) 트 픽 , (7) . (2)-(5) core size, ( 통한 line fitting ) , . .
contour contour . lower countour , upper contour , sliding window 통해 (8), (9), (10) . ( ) temporal context 20 . ( 10, 10)
decorrelate하 LDA 통합한. (cf. 6) 환하 화한. 환 A 해 훈 (within class scatter) 행 Sw (between class scatter) 행 Sb . (scatter) 행 HMM . 행하 . LDA .



𝜇𝑖 𝐴𝑇𝜓𝑖 𝑆𝑤−1𝑆𝑏 . reduction(?) 큰 m m . LDA 환한 HMM 훈 .

6 통
한 HMM , 훈, 해 ESMERALDA 5 . HMM 512 Gaussian mixtures with diagonal covariance matrice( 2048) 포함하 semi-continuous . 52 , 10 , 12 호, Baum-Welch . conbined model . Viterbi beam- search .
13 한, / . , . HMM . HMM , . HMM 택하 . 훈 해 해 . . soft vector quantization .
, 통 통합 . x 통합 통 화하 𝑤̂ .

𝑤̂ = 𝑎𝑟𝑔𝑚𝑎𝑥𝑤𝑃(𝑋)𝑃(𝑋|𝑊)

P(W) w 형 확 P(X|W) x 한 확. absolute discounting backing-off for smoothing of probability distribution . (cf. e.g. 3)

7
, , 행했. 표 1 . , 3 , 4 . IAM a..d . 표 2 휘-free .
Senior 282, 141 행했, perplexity 15.3. 13.3% 택하 12.1% . LDA 환한 12 . (표 2) 28.5%, 1.5k 10.5%. literature( ) , 훈 . 17 28.3%, 84.1%, 1.3k 16.5%. 15 6.6%, 휘 free 41.1%. 9 15.0%.
IAM 합 c03 440, 109 . . LDA(12 ) 14.2% 형( 6) 13.3% . 택한 11.1% ( 합 perplexity 12.0). 39.0%, 421( 포함) 포함한 13.9% 11 20.5% .
행했. IAM a- f( 250) , 훈 4321( a-d), 1097( e-f) . 31.3%. . 31.3% 3 10 험했 (34.8%) . 하 LDA 환한 29.1% . 통합하 22.2% ( perplexity 12.0). 60.6% .

8
, , 행해. . , / LDA 행한 .
, .

9
트 Fi799/1 German Research Foundation(DFG) .
Bern Institute of Informatics and Applied Mathematics, Horst Bunke Urs-Viktor Marti . IAM10 .


1
2
...
18

Linear Algebra and Its Applications (4th ed.) by David C. Lay

2 ~ 2.3

Valid XHTML 1.0! Valid CSS! powered by MoniWiki
last modified 2021-02-07 05:28:55
Processing time 0.0681 sec