Recognition of Hand Written English Characters

& Numerals

CS771Machine Learning : Tools, Techniques &

Application

Gaurav Krishna Y9227224Harshit Maheshwari 10290Pulkit Jain 10543Sayantan Marik 13111057

Preprocessing Feature Extraction Classification Techniques Brief Discussions on Results

◦ Parameter Selection◦ Comparative Results for Different Techniques

Final Results

Outline

We have 55 examples for each of the 62 classes

Problems Handled and Preprocessing Steps◦ Varied Size of characters Resized the characters in images of size 32X32◦ ILL centered

Centered the Images◦ Varied thickness of strokes

Thinning ( Done only for 13 Features) These steps were done in MATLAB

Preprocessing

Following Features are Considered SET 1*

◦ Haralick Texture Features◦ Zoning Feature◦ Eccentricity◦ Raw Moment ◦ Covariance

SET 2◦ Contour Feature◦ Histogram Feature◦ 13 Point Feature◦ Holes Feature

We used Java( jFeatureLib Library) to extract these features

* In graph plotting, when we say all feature we mean set 1

Feature Extraction

Random Forest Classifier Neural Network

◦ Single Hidden Layer◦ Double Hidden Layer

SVM Classifier ( Using SMO Algorithm) K-Nearest Neighbour

Classification Techniques Used

BRIEF DISCUSSION OF RESULTS

Effect of Scaling the Images

Used features is image pixel values

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 160

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Series1Series2

Determining Zoning Parameter

Zoning, Haralick Features and Eccentricity without Thinning

Deciding SVM Complexity Parameter

2 3 4 5 6 777

77.2

77.4

77.6

77.8

78

78.2

78.4

78.6

78.8

79

78.2991

78.8856 78.8856

78.2991

77.7126 77.7126

Effect of SVM Complexity Paramter

SVM Classification on Individual Features

Haralick Histogram Raw moments + Covariance

Eccentricity Holes0

2

4

6

8

10

12

Series1

Haralick 9.6774

Histogram 1.7595

Raw moments + Covariance 7.0381

Eccentricity 5.8651

Holes 2.346

Using Different Feature Sets on SVM

Holes + Zoning

Holes+ Zoning + Haralick

Contour Contour + Haral-

ick

Contour + Zoning

His-togram

His-togram + Zoning + Thinning

13 Point Feature + Thinning

Zoning Haralick + Zoning

Haralick + Zoning

+ Raw moment

+ Co-variance

Accu-racy

76.8328 78.0059 72.7273 74.1935 78.0059 34.8974 69.2082 60.4106 75.5681 78.0059 77.1261

51525354555657585

Accuracy

Using all the extracted features with SVM

Pixel values+SVM+

without thinning

Using 64 Zones 64 zones + Im-age in 32x32

pixels

SVM with 64 zones

Neural Network with 64 zones

Accuracy 74.4868 76.8328 62.7566 72.7273 61.8768

5

15

25

35

45

55

65

75

85

Accuracy

Accura

cy

Features: Haralick, Eccentricity, Zoning

Nearest Neighbors in K-NN

Single Hidden Layer Doble Hidden Layer + Feature Set 1

Contour+Zoning +Single Hidden

Accuracy 71.85 57.478 70.3812

5

15

25

35

45

55

65

75

Accuracy

Using Neural NetworkFeature Set 1 {Zoning, Haralick, Eccentricity, Raw Moments}

Using Random Forest Classifier

Zoning Contour feature + Zoning62

64

66

68

70

72

74

76

78

80

68.0352

77.7126

Random Forest Classifiers

We divided the data into 90% Training Set and 10% Test Set.

We got the accuracy of 78.8% (Using SVM on Haralick, Zoning (8X8), Eccentricity)

Ten fold cross validation gives the accuracy of 77.39%

Training error of 5% obtained.◦ Training and Testing on whole dataset gave 95%

accuracy.

Final Results

THE MNIST DATABASE of handwritten digitshttp://yann.lecun.com/exdb/mnist/

The Chars74K dataset Character Recognition in Natural Imageshttp://www.ee.surrey.ac.uk/CVSSP/demos/chars74k/

Handwritten Character Recognition using Neural Networkshttp://home.iitk.ac.in/ sunithb/NN.pdf

References

Thanks

END