COMPARISON OF IMAGE ANALYSIS FOR THAI HANDWRITTEN CHARACTER RECOGNITION Olarik Surinta, chatklaw Jareanpon Department of Management Information System.

COMPARISON OF IMAGE ANALYSIS FOR THAI HANDWRITTEN CHARACTER

RECOGNITION

Olarik Surinta, chatklaw JareanponDepartment of Management Information System

Faculty of InformaticsMahasarakham University

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

INTRODUCTION

• This article presents a comparison of image analysis for Thai handwritten character recognition.

• There are three steps of our methodology, – Data Pre-processing Method– Training Method– Pattern Recognition Method

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

INTRODUCTION

• Data Preprocessing Method – Data preprocessing method is

based on feature extraction to find the edge of character-image (described by a group of Fourier Descriptors: FD).

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

INTRODUCTION

• Training Method– Training method is based on

Back-Propagation Neural Network and Robust C-Prototype technique to processed Thai Handwritten character images to identify the centroid of each character, resulting in 44 prototypes.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

INTRODUCTION

• There are 44 Thai characters:

ก ข ฃ ค ฅ ฆ ง จ ฉ ช ซ ฌญ

ฎ ฏ ฐ ฑ ฒ ณ ด ต ถ ท ธ น บ ป ผ ฝ พ ฟ ภ ม ย ร ล ว ศ ษ ส ห ฬ อ ฮ

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

INTRODUCTION

• Pattern Recognition Method– Pattern recognition method is

based on Back-Propagation Neural Network and Robust C-Prototype technique to identify the unknown image (Thai handwritten character).

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

DATA PREPROCESSING METHOD

• Character-images are images of Thai handwritten characters.

• The output will be stored in the term of digital data by scanning. One bitmap file with gray scale pattern and 256 levels specifics one character.

Figure 1 A prototype character-image.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Binarization

• Binarization converts gray-level image to black-white image, and to extracting the object component from background, this scheme will check on every point of pixel.

Figure 2 The example of binarization scheme.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Binarization

(A) (B)

(C)

Figure 3 The diagram of extracting the object from the background component in the image.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Edge Detection

• Edge detection is one of an important image processing phases.

• This paper uses chain code technique to detect the image’s edge. The direction has been classified by 8 categories:

Figure 4 Chain code with 8 directions.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Edge Detection

• Once the edge of image has discovered, shown in figure 3, the process needs to find the character line.

• The coordinate is represented by complex number as the formula:

kkk iyxu

kk yx ,

Figure 5 coordinate represented in character image.

kk yx ,

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Fourier Descriptors

• Fourier Features used to describe the edge of the object works by identify coordinate ; K = 0, 1, …, N-1 where N is any other area in the image.

• All point will be represents as complex number.

• Therefore, the DFT can be derived as below:

kk yx ,

kk yx ,

lkNjuf

N

kkl

2exp

1

0

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Fourier Descriptors

• From the above formula, coefficient vector will be automatically calculated.

• This vector fits as 1 dimension with the size of 1x10 or 1xn

Figure 6 Fourier Descriptors of Image.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

TRAINING METHOD

• Training method is based on – Back-Propagation Neural

Network and – Robust C-Prototype technique

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Robust C-Prototype

• RCP can be determined in grouping phase in order to estimate C-Prototypes spontaneously, utilizing loss function and square distance to reduce some noise.

• The definition can be expressed as:

2

1 1

;, iji

mC

i

N

jijF duXUBJ

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Robust C-Prototype

• The diagram of solving the problem by RCP is shown in figure 7

Figure 7 RCP algorithm.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Back-Propagation Neural Network

• Back propagation neural network use a learning algorithm use in which an error moves from the output layer to the input one.

Figure 8 The Neural Network Layer.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Back-Propagation Neural Network

• Back-Propagation neural networks consists of several neuron layers, each neuron of layer being connected to each neuron of layer.

• In the training phase, the correct class for each record is known (this is termed supervised training),

• and the output nodes can therefore be assigned "correct" values -- "1" for the node corresponding to the correct class, and "0" for the others.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Back-Propagation Neural Network

• These error terms are then used to adjust the weights in the hidden layers so that, hopefully, the next time around the output values will be closer to the "correct" values.

• The network processes the records in the training data one at a time, using the weights and functions in the hidden layers, and then compares the resulting outputs against the desired outputs.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Back-Propagation Neural Network

• Errors are then propagated back through the system, causing the system to adjust the weights for application to the next record to be processed.

• During the training of a network the same set of data is processed many times as the connection weights are continually refined.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

EXPERIMENTAL RESULT

• The data in this experiment have 2 sets. – The first set is “the learning set” contai

ning 4,400 characters. – The second set is the “test set” contai

ning 440 characters.

• All data is Thai handwritten and generated by 100 persons.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

EXPERIMENTAL RESULT

Topic Robust C-Prototype

Back-Propagation neural network

Time of learning 1.5 Hour 2.45 Hour

Accuracy test on “Test set”

91.5% 88%

Comparison between Robust C-Prototype and Back-propagation neural network

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

CONCLUSION

• This paper looks at Thai handwritten character recognition. We compared Robust C-Prototype and Back-Propagation neural networks.

• The experimental results of both methods have accuracy more than 85%.

ICIIP 2006

Home

Introduction

Data Pre- Processing Method

Training Method

Conclusion

Experimental Result

Figure 9 The character images the adjustment scheme.