RENT A THINKER
Home My Page Chat Tik-Toke
2        Hand Written Character Recognition Using Neural Network

2.1       Abstract
2.2       Introduction  

2.3       Design Approach  

2.4       Character Separation Algorithm   

2.5       Character Recognition  

2.6       Results 

2.7       Reference 
2.8       Published  
  

2.1       Abstract

 Hand written character recognition is one of the focused area of research in the field of Artificial intelligence. It involves image digitisation, line separation, character separation, image enhancement, data compression, segmentation, image recognition and contextual error correction. This paper describes a neural network based technique to recognize isolated hand written characters. This method involves hierarchical feature extraction and classification.

 A neural network based algorithm is developed using Visual C++ to recognize hand written characters. The network is based on Multi Layered Feed Forward type network with 64 binary inputs and 26 binary outputs. The system consists of other modules like- page scanner, line separator, character separator, character normalizer etc. The network software performs off-line training and character recognition. This design is only for uppercase English character set.  

2.2       Introduction

The weakest link between the computers and its user is the inputting systems; in most cases is the keyboard. There are   various ways, where computers   directly take input   from the human information system like- optical character recognition (OCR), speech recognition, symbolic (Icons, windows) interactive communication etc. Such systems are difficult to design and do not provide complete error free operation. General solution for speech or character recognition is very complex. Even fastest computers would need much larger computational time compared to human response time to perform   the same job. To get working solution for the above class of problems, domain specific solutions are much more efficient.

In hand written character recognition problem, the domain is reduced to a small subset of characters of limited number of written characters using specific style. This subset is then further classified to smaller subsets, where each group represents a character. The character space is visualized as space confined by non-linear decision boundaries separating the characters from each other. Any overlap in decision boundary causes erroneous recognition. Hence the design goal is to make as crisp boundary as possible in the domain space so that ambiguity of the character subset recognition is reduced to minimum.

2.3       Design Approach

As described above, the domain space in case of digitized hand written text could be in vector or raster graphics format. In case of on-line text entry (using mouse, light pen or graphics tablet) each individual strokes (vector) is recorded. and in case of text data on paper, raster scanning method is used for digitization using optical scanner or video camera. Raster to vector conversion is very complex process. Hence we choose the raster format,   as standard input for both on-line and off-line texts.  A scanner of 600 dpi  gives approximately 64 x 64 pixel resolution for an average hand written character.  Figure-1 shows the block diagram of the proposed system.

 Figure-1.  The block diagram of a Hand Written Character Recognition System using Neural Network based feature extraction and feature classification.

 The size and shape of a hand written character may vary considerably in a given text.  It is necessary to normalize both size and shape of a character before presenting it to an OCR engine.  Size normalization needs contraction and/ or expansion with minimum distortion in aspect ratio, so that the character fits to a fixed template (e.g. 24x24).

 The size-normalized data is segmented in small zones (e.g. 8x8) to examine the content to codify and quantify in terms of shape parameters. This is achieved by vector quantization technique as shown in figure-2. Each segment is represented by a small set of parameters (shape vector). This process further reduces the data size approximately to one tenth and normalizes the shapes.

In the next step all the shape vectors corresponding to a character are assembled back. The feature vector set is then classified using another multilayered neural network (72 - 64 - 26). There are 26 outputs of this network each represents a specific character that may be written in different shapes or sizes. The model of vector quantization and classification is based on large data set and needs to be computed off-line using supervised training.  

2.4       Character Separation Algorithm

 A statistical sampling technique is used to increase the speed of   the character separation   from hand written script.  In this method random samples are taken from middle of each side of the paper to locate the limits of the script boundary. Similarly, the line and character boundary are detected by searching horizontal blank line for line separation and vertical blank space between consecutive characters. This method is found 10 - 15 times faster than the continuous raster scan method.  However, being statistical in nature, the method at times misses the gaps and combines two characters together. An alternative method is used   to scan alternate pixels and scan lines. This method is four times faster and found error free. Figure-3 is showing character isolation and boundary detection result.  The boundary separation data is stored in a separate file and is presented to the character normalization module.  

  Figure-2. Feature extraction from a hand written character segment. In above diagram 64 output neurons are trained to reproduce 64 input neurons (8x8 image segment) through 8 hidden neurons. The hidden node outputs are used to further classify the character. 

2.5       Character Recognition

 The character recognition involves comparing of target character’s attributes with the set of stored templates. The best match, if found above a pre-defined threshold, is declared as recognized to corresponding template. For contextual recognition it is necessary to store a list of top matches. Simple spell check algorithm is used to select the most appropriate character from the list. The selected character may not be always at the top of the list.  Also when the character is outside the domain of stored templates, the best match is found below the acceptable threshold value. It is desirable to choose optimum set of attributes for the best results. The total score of match is computed by- 

     M=S at x as 

Where at   is   attribute of the template
 
and  as  is  the attribute of the sample.

  

 

 

 

 

 

 

 

 

Fig-3 A sample run of Hand Written Character recognizer package developed by author.

 In the hand written characters the attributes could be visualized as different shape properties associated with a character like for e.g. curvature, straight line, corner etc. To extract dominant shape parameters from a large set of characters, the character field is divided into small segments (8 x 8). Each segment data is presented to a sand-clock shaped neural network having equal number of input and output nodes and fewer hidden nodes (see figure 2). When the network reproduces the input at the output (using back propagation learning algorithm), the hidden neuron outputs the dominant features. To recognize a sample character, it is divided into small zones  (3 x 3) and the dominant attribute corresponding to each zone is measured. The classification of this set is tested with another feed forward type neural network (see figure-1), which is previously trained with the full set of templates. 

2.6       Results 

 A prototype handwritten character recognition package is developed   using Visual C++ to recognize handwritten upper case English letters. The system uses mouse as input device. The input text is read from the graphic screen directly. Figure 3 shows a sample output of the package. 

2.7       Reference

1.      Himanshu S. Mazumdar, Leena Rawal, “A Learning Algorithm for Self Organized Multilayered Neural Network” CSI Communication, pp.5-6(May 1996) 

2.      Leena P. Rawal, Himanshu S. Mazumdar, “A DSP based Low Precision Algorithm for Neural Network using Dynamic Neuron Activation Function” CSI Communication, p.15-18 (April 1996) 

3.      Leena P. Rawal, Himanshu S. Mazumdar, “A Neural Network Tool Box using C++” CSI Communication, pp.22-25 (April 1995) 

4.      Hetcht-Nielsen, "Theory of the back propagation neural network, “in Proc. Int. Joint Conf. Neural Networks, vol.1, pp.593-611. New York: IEEE Press, June 1989. 

5.      Lippmann R.P., An introduction to computing with Neural Nets.IEEE ASSP Magazine, pp4-22, April 1987. 

6.      Rumelhart and J.L. McClelland Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Cambridge MA: MIT Press, vol. 1,1986.

 

2.8  Published:
Himanshu S. Mazumdar., “Hand Written Character Recognition using Neural Network " CSI Communication, p.16-18 (Feb. 1998)