72

Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 2: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 3: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Noise cleaning and Binarization

Skew Correction

Text & Graphics Segmentation

Line & Word Segmentation

Parsing (CC Analysis)

Feature Extraction

Classification Converter & Post-processing

Document Reconstruction

Output: Text/ unicode

Input: Document image

Page 4: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 5: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 6: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 7: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 8: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 9: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

×

Page 10: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Input

CC Analysis

Convert to symbols

Reorder symbols

Render the word image

CC Analysis

Labels the CCs

33 51 122 52 113 107

DP based

Matching

to align

R and W

MAP FILE

RULES FILE

R

Page 11: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 12: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 13: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

CC Analysis Label 37 55 107 57 37 58 43 63 14

Feature Extraction

Page 14: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 15: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Page 16: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 17: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 18: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 19: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 20: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

•••••••••

IIIT Hyderabad

Page 21: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 22: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 23: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Feature Dim Classifiers

MLP KNN ANN SVM-1 SVM-2 NB DTC

C.M 20 12.04 4.16 5.86 10.04 9.19 11.93 5.57

DFT 16 8.35 8.96 9.35 7.88 7.86 15.33 13.85

DCT 16 5.43 5.11 5.92 5.25 5.24 8.96 7.89

ZM 47 1.30 1.98 2.34 1.24 1.23 3.99 8.04

PCA 350 1.04 1.14 2.39 0.37 0.35 4.83 5.97

LDA 350 0.55 0.52 1.04 0.35 0.34 3.20 4.77

RP 350 0.33 0.50 0.74 0.34 0.34 3.12 8.04

DT 400 1.94 1.27 1.98 1.84 1.84 4.28 2.20

IMG 400 0.32 0.56 0.78 0.32 0.31 1.22 2.45

Error rate using CNN : 0.93

IIIT Hyderabad

Error rates on Malayalam dataset.

Page 24: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 25: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 26: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Error rates of SVM-2 classifiers with varying number of features.

Page 27: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 28: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 29: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Accuracy of different classifiers Vs no. of classes, Feature used : LDA.

Page 30: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad`

Page 31: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 32: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Images from dataset

IIIT Hyderabad

Page 33: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Feature D-1 D-2 D-3 Blobs Cuts Shear

C.M 9.45 9.46 10.97 16.28 12.33 30.07

DFT 7.89 7.93 7.98 26.70 8.73 18.90

DCT 5.71 5.72 6.07 19.80 7.93 16.46

ZM 1.96 1.98 2.10 8.41 4.35 17.75

PCA 0.39 0.39 0.40 2.17 0.64 8.59

LDA 0.30 0.31 0.32 2.01 0.61 7.32

RP 0.48 0.67 1.04 3.61 0.71 6.75

DT 1.75 1.98 2.21 10.33 5.07 12.34

IMG 0.32 0.33 0.33 2.78 0.66 6.84

IIIT Hyderabad

Page 34: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 35: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 36: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the 5th font. S1 : Dataset without degradation, S2: Dataset with degradation.

Page 37: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 38: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Features Telugu (350 class) English (72 class)

20X20 40X40 20X20 40X40

C.M 20.78 12.32 7.25 6.48

DFT 8.45 5.48 2.04 1.12

DCT 9.67 2.71 2.14 1.04

ZM 15.71 6.71 5.37 3.31

PCA 4.62 2.93 0.86 0.46

LDA 2.56 1.67 0.29 0.23

RP 2.49 1.66 0.28 0.23

DT 3.48 3.17 0.98 0.87

IMG 3.18 2.84 0.28 0.23

IIIT Hyderabad

Page 39: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

IIIT Hyderabad

Page 40: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 41: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

1,5

2,5 1,4

3,5 2,4 1,3

4,5 3,4 2,3 1,2

x

5 4 1 2 3

Sample x from class 4

Page 42: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 43: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 44: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Page 45: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 46: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 47: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 48: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 49: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 50: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

|C|

O)stance(C,CharEditDi (CER) DistanceEdit Character

Symbols leRecognizab of No. Total

Symbols leRecognizab and iedMisclassif of No. RateError Symbol

Unicodeof No. Total

UnicodeiedMisclassif of No. RateError Unicode

Page 51: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 52: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 53: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Wordsof No. Total

rdsCorrect Wo of No. Accuracy level Word

WordsleRecognizab of No. Total

WordsleRecognizabor Correct of No.Accuracy level Word

Page 54: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 55: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 56: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 57: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 58: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 59: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 60: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Sarada

Page 61: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Sanjayan

0.85%

Page 62: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Thiruttu

0.85%

Page 63: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 64: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 65: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

••

• •

Page 66: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 67: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 68: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Page 69: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the

Page 70: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 71: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the
Page 72: Recognition of Malayalam Documents IIIT Hyderabad › ... › neebaPresentation2010.pdf · IIIT Hyderabad Accuracies of SVM-2 classifier when trained with 4 fonts and tested on the