MLP-based Assamese Character and Numeral

3rd Indian International Conference on Artificial Intelligence (IICAI-07)

MLP-based Assamese Character and Numeral Recognition using an Innovative Hybrid Feature Set Kandarpa Kumar Sarma Department of Electronics Science Gauhati University, Guwahati-781014, India. [email protected] Abstract Neural number recognition as an extension of an Optical Character Recognition (OCR) system requires a unique feature set capturing relevant details of the input profiles. The performance of such a recognition system depends on the feature set. The work deals with the design of a combined character and numeral recognition system using a hybrid feature set applied earlier to the recognition of Assamese characters. The work also describes the use of a modified hybrid feature set derived for handwritten numeral detection using a neural network. The set aims to maximize recognition performance, improve robustness and invariance to shape and size in presence of noise.

1

Introduction

Neural networks have become a preferred method for a range of complex pattern classification problems. Neural networks like the multi layered perceptrons (MLPs) permit connectionist nonlinear computing. Non-linear computing resembles the parallelism generated by the human brain. The hidden layers of an MLP helps the network to learn complex patterns and perform their classification [1] [2] [3]. These possibilities allow the use of a MLP for the design of a combined character and numeral recognition. Considerable work has been done in Optical Character Recognition (OCR) of Indian languages. A detailed account of such work is available in [4]. Some of the previous work done in the field of numeral recognition in Indian languages include works done in Bengali [5] [6], Devnagari and English handwritten numeral recognition [7] etc. A neural network based combined character and numeral recognition system for Assamese script using MLP based classification is proposed in this work. The scheme has been tested for the best mean square error (MSE) convergence rate and classification performance using an innovative hybrid feature set already formulated [4]. The work also includes the use of multiple MLP configurations and tested with four different learning rates and two momentum constants. The idea is to determine the most suitable combination with respect to learning rates and momentum constants to be used with the MLP configuration for recognition. The present work uses a hybrid feature set formed by a varied mixture of morphological, statistical, geometrical and tomographic projection features Bhanu Prasad (editor): IICAI-07, pp. 585-600, 2007 Copyright © IICAI 2007


Figure 1: Assamese numerals used for training MLP classifier for recognition of Assamese characters and numerals. This hybrid feature set has already been used to perform recognition tasks involving characters of Assamese Script [4]. The feature set provides the best MSE convergence and highest classification and recognition rates. The present work also proposes a modified hybrid feature set used to train a neural network which exclusively performs the recognition of Assamese handwritten numerals. The set aims to maximize recognition performance, improve robustness and invariance to shape and size in presence of noise in case of numeral recognition only. The modified feature set is derived from the hybrid feature set [4] already referred. This paper is organized into the following sections. In Section 2 Assamese Numerals and its characteristics are discussed. Section 3 describes the algorithmic steps followed in the work. Details of the feature extraction process are explained in Section 4. Description of the formulation of the modified hybrid features for use with handwritten numeral recognition is given in Section 4.2. Configuration of the MLPs and training are explained in Section 4.3. Performance details of the earlier proposed hybrid feature set used for combined character and numeral recognition is included in Section 4.4. Results obtained from the newly proposed modified hybrid feature set is provided in Section 4.5. A conclusion is included in Section 5.

2

Assamese Numerals and characteristic features

Assamese is an important language in the North Eastern part of India. Many Indian languages including Assamese and Bengali have the same origin-Brahmi, hence there are certain similarities in the general shape and appearance [8] of the characters and numerals. Some characteristic features of Assamese characters are available in [4]. A sample set of Assamese characters used for the training of the MLPs for performing the combined character and numeral recognition is available in ( Figure 2 of [4] ). Assamese numerals are similar to that used by Bengali language. Isolated numerals in Assamese are characterized by the presence of zero-like, multiple zero-like, compound zero-like and straight line containing shapes, curvatures, compound shapes with curvature and lines etc ( Figure 1). Handwritten numerals can contain huge amount of variations due to differences 586


Figure 2: Handwritten Assamese numeral samples used for training in writing styles, orientation etc (Figure 2). Writer variation results in modification and alteration of shape, size, inclination and distribution of numerals. Certain writing styles make numerals touch each other. These variations provide ample of opportunity for a character recognition system to innovate methods for achieving higher rates of recognition. A neural network is well suited for such applications because of its ability to learn adaptively.

3

Algorithmic steps of the present work

The entire system maybe described by the following algorithmic steps: • Training: 1. Generation of hybrid feature vectors of all possible classes. These include two font and size variations of characters including special characters, modifiers and compound characters. Also included in the set are ten numerals in Assamese script. Total number of classes thus become 99. In this case 11 vowels, 41 consonants, 19 compound characters, 18 special characters ( Figure 2 of [4] ) and 10 numerals have been used ( Figure 1 ). 2. Designing and training the MLPs in three configurations (with error Back Propagation (BP)) with four learning rates and two momentum constants. The three different configurations are taken to ascertain the best configuration with respect to performance. 587


3. Performing classification for all available classes of characters and numerals used for training the MLPs in all the configurations. The MLP with the best MSE convergence and classification performance is used to perform recognition. 4. Performing classification for all available classes of characters and numerals used for training the neural network with noise mixed data. The convergence of the MSE and classification performance are noted. 5. Determining the best MSE convergence and classification performance with respect to feature type and MLP configuration. • Testing: 1. After pre-processing, segmentation is performed. At first, segmentation is done to obtain lines from page, next words from lines and finally letters and characters from words. A similar segmentation is carried out for the numerals given to the system as text input. 2. Then hybrid feature vectors are generated from the testing set taking two font and size variations and italic types in Assamese script. Special characters, modifiers and compound characters are also included in the testing set using the selected feature type. The numeral set taken includes multiple size variations with noise free and noise mixed forms. 3. The feature vector is given as input to the selected trained MLP. 4. If the characters belong to the proper class the output is displayed in the transformed domain and thus recognition is complete. A similar set of experiments are carried out exclusively for the handwritten numerals as well.

4

Feature Extraction

The detailed description of the feature extraction process with respect to the formation of the hybrid feature set is available in [4]. That feature set is of length 223 which makes it suitable enough for handling upto 111 classes of input by following the norm of assigning two features samples per input [9] [10]. In this work the total number of considered classes is 99. Hence, the feature set was used to perform a MLP based recognition of a combined input of Assamese characters and numerals. The results of the experiments carried out is provided below in Section. 4.4. Some modifications, however, are carried out and a new feature set formulated ( Section 4.2 ) in connection with the MLP based recognition of handwritten numerals.

4.1

Assamese Numeral Recognition using MLPs

In this work, MLPs in two configurations are trained with a set of modified hybrid features of Assamese handwritten numerals. The feature vectors are generated from the scanned input of ten handwritten Assamese numerals as shown in Figure 1. The detailed 588


feature extraction process is described in the Section 4.2. The specific MLPs are therefore trained to classify upto 10 types of inputs. The training data set is derived during the feature extraction phase from a set which consists of 1000 samples per numeral as shown in Figure 2. The entire work has two part: training and testing. During training, modified hybrid feature vectors of all ten classes of numerals in Assamese script are generated. Two different MLP are formed and trained with four learning rates and two momentum constants to ascertain the best configuration with respect to MSE convergence and classification performance. The network with the best MSE convergence and classification performance is used to perform recognition. The proposed modified hybrid feature set of length 50 gives better performance than the case when the original hybrid feature set [4] of length 223 is used directly to train the MLP classifier. During testing, first, preprocessing is done. The common steps performed as part of pre-processing include noise removal, enhancement of the input image, converting into a binary form etc. Segmentation separates each numeral from the scanned input. Then feature vectors are generated from the testing set taking several written samples so that the content includes multiple writer- induced variations. Next, the feature vector is given as input to the selected and trained MLP. If the input character belongs to the proper class, the output is displayed in the transformed domain to complete the recognition. These steps are repeated with noise mixed data too.

4.2 Modified Hybrid Feature Extraction The length of the feature vectors used for training the MLPs for handwritten numeral recognition is 50 for each of the 10 numeral classes. The feature extraction process is aimed at making the process robust enough to deal with three sizes, 8x8, 16x16 and 32x32. These size variations however may be discarded by considering a normalized size of either 16x16 or 32x32. The consequence of this measure will be a reduction in the training time but will provide no size variation to the system which is important in character recognition systems. The proposed modified set is aimed at providing better invariance to shape, size and inclination. The length of the feature vector is fixed at 50 due to the fact that this feature size gives the best MSE convergence and recognition rate as shown by Table. 1. These results have been generated using a 2−hidden layered MLP trained for 7000 epochs with Gradient descent with momentum back propagation and tested with a sample size of 1000 for each of the classes. The modified features are derived from the original hybrid feature set by extracting only those values that have a variance difference between 20 to 25 percent between adjacent samples. By following this criteria, feature vector length of 50 is derived. The effect of variation of average inter-sample variance of the feature set is seen in Table 1. The increase in the inter-sample variance of the feature sets beyond the 21% limit ( with feature length 50 ) decreases the average MSE convergence and recognition rates. It means some vital information regarding the input numerals are lost if the feature length shrinks beyond 50. Similarly, if the inter-sample variance of the feature set is decrease it leads to an increase in the feature set length. But the increase in feature set length is coupled to a reduction in the MSE convergence and recognition rate of the MLP. Increase in feature length provides the MLP more number of samples to learn. It improves learning but it helps the MLP to memorize. As a result the MLP fails to generalize. A neural network when memorizes less and generalizes more acts as a superior classifier. Also, 589


Table 1: Effect of feature vector length on MSE Convergence and Recognition rate using a 2-hidden layered MLP. Case Feature Average MSE Average Av. interLength Convergence Class wise sample varia Recognition -nces of in % feature sets in % −3 1 223 2x10 85.2 1.5 2 200 1.22x10−3 83.6 3 −4 3 170 1.72x10 77.8 6 −5 4 140 2.12x10 78.6 9 5 110 3.02x10−5 79.3 12 6 80 2.12x10−5 78.6 15 −5 7 50 0.12x10 95.6 21 −4 8 40 1.7x10 91.6 24 9 30 1.61x10−4 90.6 27 −4 10 20 1.43x10 91.6 30 Table 2: Breakup of different contributions towards the formation of the modified hybrid feature vector Contributed Quantity Vector by Length Morphological 17 50 Statistical 17 Geometrical 9 Tomographic 12 though more number of samples helps the learning of an MLP, it ushers in the curse of dimensionality. Increasing feature length decreases MLP convergence beyond the noted limit due to the famous local minima problem [1] [2] [11]. The modified features for nine numerals in Assamese can be seen in the Figure 3. The modified hybrid feature vector represents a near complete description of the profile of a numeral. A breakup of the contributions by different segments of the feature vectors is shown in Table 2. The original hybrid features as reported in [4] and [12] is formed by contributions of morphological, statistical, geometrical and tomographic projection features. The percentage wise breakdown of these contributions are shown in Table 3. The percentage wise breakdown of different feature components of the original hybrid feature set as shown in the Table 3 nearly matches the breakdown of contributions (Table 2) made by morphological, statistical, geometrical and tomographic projection features towards the formation of the modified feature set used in this work. The contribution of the morphological and statistical features to the hybrid set is maximum. Morphological features represent the widest variation of feature samples linked with boundaries, shape, size and connectivity of a handwritten numeral. This variation conveys certain unique information through which representation of a numeral’s profile becomes invariant to shape, size, connectivity, discontinuity etc. As a result the feature set incorporates all the components through 590


200

200

200

100

100

100

0

0

0

−100

0

20

40

60

−100

0

20

40

60

−100

150

150

150

100

100

100

50

50

50

0

0

0

−50

0

20

40

60

−50

0

20

40

60

−50

200

200

200

100

100

100

0

0

0

−100

0

20

40

60

−100

0

20

40

60

−100

0

20

40

60

0

20

40

60

0

20

40

60

Figure 3: Modified hybrid features for 9 handwritten numerals in Assamese.

Table 3: Breakup of different contributions towards the formation of the hybrid feature vector of length 223 Contributed Quantity Contribution by in % Morphological 77 35 Statistical 74 34 Geometrical 40 18 Tomographic 32 13

591


which it is better placed to handle writer-generated variations. Also the subtle variations in the morphological profile between individual characters can be captured proficiently by the morphological features. Equally important is the statistical feature because it has also made a considerable contribution towards the formation of the modified hybrid feature set. Since the statistical features provide quantitative description of the shape of the numeral image boundaries, this contribution aids the feature set in tackling noise mixed inputs. This way statistical features received greater importance and made the highest contribution along with the morphological features in forming the hybrid feature set. Also statistical components have considerable amount of principal components which have less correlation among the components of the feature set and minimize the effects of dimensionality [11]. Tomographic projections are required to obtain multi angle views of image intensity variations. Finally, the geometrical features has made the least number of contributions which is equally important like the other feature contributions. The geometrical features represent information related to the physical profile of a character. Assamese numerals show geometrical variations due to the presence of corners, cusps, horizontal and vertical distributions, zeros, multiple zeros etc. These geometrical profiles at times show certain similarities but have minor variations as well which differentiate a numeral from the other. The geometrical feature set is well suited to extract these variations in the physical profile of the numerals.

4.3 Configuration of the MLPs and Training One hidden layer of an MLP is sufficient to approximate the mapping of any continuous function [2], [3] . However, with many class problems like the present one, multiple hidden layered MLPs provide better MSE convergence, classification performance and subsequently higher recognition rates; but if the number of hidden layers increases though non-linearity is increased ( meaning increase in computational capability ), the MLPs tend to memorize and reject generalization in classification [2]. Several configurations of the MLP are utilized for training. The sizes of the hidden layers are selected without following any logical reasoning; all are randomly selected values [3]; but the size of the input layer depends upon the length of the feature vector and the length of the output layer reflects the number of classes. The configuration of the MLPs used for the combined Character and Numeral Recognition System maybe shown in a tabular form as in Table 4 where P refers to the size of the feature vector. The training of the MLPs is performed with varying learning rates like 0.01, 0.06, 0.1, 0.6. As seen from the values the four learning rates have been so selected that the alternate ones are multiples of 10. As a momentum term in the weight update equation accelerates the convergence of the MSE to the desired performance goal, two momentum rates have been utilized. The two momentum rates are 0.4 and 0.6. Larger the momentum rate, larger is the convergence but there is again a possibility of the MSE convergence showing oscillation around a local minimum[1]. Similarly, for the exclusive numeral recognition system, for ten classes, a single hidden layered-MLP is sufficient. But handwritten numeral recognition is a task with considerable complexity which requires multiple hidden layered-MLP to perform the classification and recognition. It allows the MLP to learn the writer-induced variations of the inputs and deal with related variations. The configuration of the MLPs used in this case may be shown in a tabular form as in Table 5.

592


Table 4: Configuration of the MLPs for combined character and numeral recognition Cases Hidden Layers Neurons/hidden layer P-feature vector length 1 One 1.5*P 2 Two 1.5*P, 0.89*P 3 Three 1.5*P, 0.89*P 0.6 *P Table 5: Configuration of the MLPs for handwritten numeral recognition Cases Hidden Layers Neurons / hidden layer P-feature vector length 1 Two 1.5*P, 0.89*P 2 Three 1.5*P, 0.89*P 0.6 *P

4.4

Performance of MLP Configurations using Hybrid Feature Vector applied for combined character and numeral rcognition:

Classification results are calculated during the training phase by noting the number of correct class wise placements by the MLPs as compared to the number of inputs. For example, if the input to an MLP is 200 and the number of correct classification is 186 then the successful classification rate is taken to be 93%. Hybrid features provide classification rates (between 94% and 95.7%). Hybrid features show higher levels of uncorrelatedness in the data. This uncorrelatedness helps the MLP during training by preventing it from getting stuck in some local minima. As a result the MLP has greater opportunity to adapt and update its connectionist weights between the layers. It develops higher degree of generalization but there is again a danger of overtraining which helps the MLP more to memorize. The desired convergence level therefore needs to be carefully fixed during training. All measures must be taken to make the uncorrelatedness as low as possible. Hybrid features appears to be robust as was previously reported in [4] this time in classifying both characters and numerals of Assamese language . Classification results of 99 classes of input generated with 10 sets of test data each containing a input text of 250 characters and numerals is shown in Table 6 Recognition results are calculated during the testing of correct class wise placements by the MLPs as compared to the number of inputs. For example, if the input to an MLP is 200 and the number of correct recognition is 186 then the successful recognition rate is taken to be 93%. As hybrid features provide the best classification performance, these are used for recognition of test data. Tests have been carried out with several sets (each of 250 characters) having size and font variations. Italic characters included adds more variety to the input set during recognition. Noise free and noise mixed inputs have also been included to find out the robustness of the MLP. Noise, with variations of 10dB, 30dB, 50dB, 70dB and 90dB, is mixed with the input text data from where individual characters 593


Table 6: Average Classification with hybrid features for combined character and numeral recognition Cases Neurons/Layer Classification with learning rate(lr) %

1 2 3

1st 1.5 ∗ P 1.5 ∗ P 1.5 ∗ P

2nd x 0.89 ∗ P 0.89 ∗ P

3rd x x 0.6 ∗ P

lr = 0.01 94.0 94.6 95.2

lr = 0.06 lr = 0.1 94.4 93.2 95.3 95.3 94.6 95.1

lr = 0.6 94.2 95.7 95.2

Table 7: Average Class wise Recognition rate(2-hidden layered MLP) Case Trained Tested Average with with Class wise feature Data Recognition vectors that Rate in that are are percent 1 Noise Free Noise free 95.2 2 Noise Free Noise Mixed 83.6 3 Noise Mixed Noise Free 96.3 4 Noise Mixed Noise Mixed 85.7 are segmented out before passing the characters to the feature extraction process. The recognition rates obtained with noise free inputs and noise mixed inputs in case of two MLP configurations trained by noise free and noise mixed feature vectors are shown in Tables 7 and 8. The tests have been carried out for two and three hidden layered MLPs. • Two hidden layered MLP: The best recognition rates are obtained for Case 3 where the MLP was trained with noise mixed feature vectors and tested with noise free data. The lower recognition rate shown by Case 2 with the MLP trained with noise free feature vectors and tested with noise mixed data is primarily due to the problems encountered in the segmentation process. The recognition rate is directly effected by the performance of the segmentation stage(Table 7). • Three hidden layered MLP: Table 8: Average Class wise Recognition rate(3-hidden layered MLP) Case Trained Tested Average with with Class wise feature Data Recognition vectors that Rate in that are are percent 1 Noise Free Noise free 94.8 2 Noise Free Noise Mixed 82.6 3 Noise Mixed Noise Free 96.4 4 Noise Mixed Noise Mixed 87.8

594


Figure 4: Segmented output of scanned input of numerals The best recognition rates are obtained for Case 3 where the MLP was trained with noised mixed feature vectors and tested with noise free data(Table 8). This is due to the fact that noise mixed training data has less correlatedness as a result of which training is better performed. It means that the MSE convergence is better as the convergence limit decreases to a lower value. With better MSE convergence, the MLP is able to discriminate between the decision boundaries with finer detail as a result of which recognition improves. The recognition rate shown by Case 2 with the MLP trained with noise free feature vectors and tested with noise mixed data is lower.

4.5 Performance of MLP Configurations using Modified Hybrid Feature Vector used for handwritten numeral recognition In this work modified hybrid features have been used for training two different configurations of MLPs with four different learning rates and two momentum constants. During testing, after pre-processing the scanned inputs of the numerals are segmented. Figure 4 shows an output of the segmentation stage. Average recognition performance of a two hidden layered MLP trained for 7000 epochs with Gradient descent with momentum back propagation and tested with a sample size of 1000 for each of the classes is shown in Figure 5. The average recognition rate achieved in the work is around 96%. However, the rate of false recognition is also considerable. In some cases it is around the 15 − 20% range which is undesirable. It maybe a case of overtraining of the neural network. Another possible reason maybe excess or additional feature samples used per numeral during training leading to dimensionality problems. The second possible reason indicates the necessity of application of data pruning methods in extracting the features for training the neural network. Noise free and noise mixed inputs have also been included to extend the range of testing of the recognition ability of the MLP. Noise, with variations of 10dB, 30dB, 50dB, 70dB and 90dB is mixed with the input text data from where individual numerals are segmented out before passing the numerals to the feature extraction process. The recognition rates obtained with noise free inputs and noise mixed inputs in case of two MLP configurations trained by noise free and noise mixed feature vectors are shown in Tables 9 and 8. The 595


Figure 5: Recognition results Table 9: Average Class wise Recognition rate (2-hidden layered MLP) for handwritten numeral recognition Case Trained Tested Average with with Class wise feature Data Recognition vectors that Rate in that are are percent 1 Noise Free Noise free 96.3 2 Noise Free Noise Mixed 84.8 3 Noise Mixed Noise Free 97.2 4 Noise Mixed Noise Mixed 89.7 tests have been carried out for two and three hidden layered MLPs. • Two hidden layered MLP: The best recognition rates are obtained for Case 3 where the MLP was trained with noise mixed feature vectors and tested with noise free data. The lower recognition rate shown by Case 2 is with an MLP trained with noise free feature vectors and tested with noise mixed data. • Three hidden layered MLP: The best recognition rates are obtained for Case 3 where the MLP was trained with noised mixed feature vectors and tested with noise free data (Table 10). This is due to the fact that noise mixed training data has less correlation among feature samples as a result of which training improves. It means that the MSE convergence is better as the convergence limit decrease to a lower value. With better MSE convergence, the MLP is able to discriminate between the decision boundaries with finer detail as a result of which recognition improves. The recognition rate shown by Case 2 596


Table 10: Average Class wise Recognition rate (3-hidden layered MLP) for handwritten numeral recognition Case Trained Tested Average with with Class wise feature Data Recognition vectors that Rate in that are are percent 1 Noise Free Noise free 96.1 2 Noise Free Noise Mixed 84.0 3 Noise Mixed Noise Free 97.1 4 Noise Mixed Noise Mixed 89.2 Table 11: MSE, Classification and Recognition Rate variation for varying SNR values for a 2-hidden layered MLP with momentum constant, µ = 0.6 used for handwritten numeral recognition SNR MSE Classification Recognitionin after Rate (%) dB 5000 Rate (%) epochs 10 0.00051 81.3 86.4 30 0.00042 86.6 88.3 50 0.00043 88.1 89.8 70 0.00040 93.2 95.1 90 0.00036 96.6 97.2 with the MLP trained with noise free feature vectors and tested with noise mixed data is lower. Table 11 shows the effect of noise on the classification and recognition process obtained with the variation of MSE, Classification and Recognition Rates for varying signal to noise ratio (SNR). The classification rate is an average value obtained with two different MLP configurations trained with four different learning rates and two momentum constants. The recognition rate is obtained from successive tests carried out with 1000 different samples of the 10 numerals in case of a two hidden layered MLP. The set includes two size variations, noise free and noise mixed inputs. The advantages offered by the modified hybrid feature set is obvious.

5

Conclusion

Neural classifiers demonstrate adaptive learning of complex patterns like character profiles even with font, size and inclination variations irrespective of conditions with and without noise. To perform such tasks, a neural classifier requires a feature set robust enough to capture all relevant information regarding shape, size, inclination etc variations. An innovative hybrid feature set earlier formulated was found to be successful in dealing with characters showing such variations. That feature set while dealing with Assamese 597


−4

5.2

MSE vs SNR plot

x 10

5

4.8

MSE

4.6

4.4

4.2

4

3.8

3.6 10

20

30

40

50 SNR

60

70

80

90

Figure 6: Variation of MSE with SNR for handwritten numeral recognition

Classification Rate in % vs SNR plot 98

96

94

Classification Rate

92

90

88

86

84

82

80 10

20

30

40

50 SNR

60

70

80

90

Figure 7: Variation of Classification rate with SNR for handwritten numeral recognition

598


Recognition Rate in % vs SNR plot 98

96

Recognition Rate in %

94

92

90

88

86 10

20

30

40

50 SNR

60

70

80

90

Figure 8: Variation of recognition rate with SNR for handwritten numeral recognition characters was found to be capable enough to capture all related details of a character in presence of noise as well. The capability of that feature set was extended in this work by applying it for recognition of characters and numerals of Assamese language. The success rate achieved is healthy. Such a feature set formed by a varied mixture of morphological, statistical, geometrical and tomographic projection feature samples and with subtle modifications is robust enough in dealing with another complex recognition task like handwritten numeral recognition. The variations observed in shape, size, orientation, inclination, distribution etc. due to the variations in writing styles are handled well by a properly trained neural network. But the difficulty arises during the training of the neural network. It is with respect to the number of training epochs a neural network must be subjected to before halting. Also, a problem arises in case of fixing the MSE convergence limit. It at times leads to overtraining which results in incorrect recognition. An adaptive training process linked with an optimal recognition rate can be a solution. Another problem is related to excessive data used as training samples. Future directions can be with respect to the formulation of an adaptive training methodology linked to the recognition rates to be achieved and the design of a smart feature extractor. As a solution to the first problem, hybrid classifiers can be used which can improve the performance further and reduce the miss hits. Another closely related aspect is the design of a feature extractor which can lead to data pruning during the formation of the feature set for training the neural network. Data pruning of the feature set can reduce the ill effects of dimensionality. These modifications can help the system evolve into an improved handwritten numeral recognizer which can also be integrated to an Assamese OCR system.

References [1] S. Haykin. Neural Networks. Pearson Education, 2003. 599


[2] C. Bishop. Neural Networks for Pattern Recognition, First Indian Edition. Oxford University Press, 2003. [3] B. Krose and P. Smagt, An Introduction to Neural Networks,Eight Edition. University of Amsterdam, 1996. [4] Kandarpa Kumar Sarma, Prabin Kumar Bora and Chitralekha Mahnata, “Innovative Feature Set for Multi Layered Perceptron (MLP) Based Assamese Character Recogntion”. Proceedings of 2nd Indian International Conference on Artificial Intelligence (IICAI-2005), pp. 473 − 491, 2005. [5] U. Pal and B. B. Chaudhuri. “Automatic Recognition of Unconstrained Off-Line Bangla Handwritten Numerals ”. T. Tan, Y. Shi, and W. Gao (Eds.): LNCS 1948 , ICMI 2000: 371 − 378, 2000. [6] U. Bhattacharya, T. K. Das, A. Datta, S. K. Pauri and B. B. Chaudhuri. “A Hybrid Scheme for Handprinted Numeral Recognition Based on a Self-Organizing Network and MLP Classiifers”. IJPRAI: Vol. 16 , No. 7(2000): 845 − 864. [7] G. S Lehal and Nivedan Bhatt. “A Recognition System for Devnagri and English Handwritten Numerals ”. T. Tan, Y. Shi, and W. Gao (Eds.): LNCS 1948 , ICMI 2000: 442 − 449, 2000. [8] B. B. Chaudhuri and U. Pal. “A Complete Printed Bangla OCR System”. Pattern Recognition, 31(5): 531 − −549, 1998. [9] U. Pal and B. Choudhury, “Automatic Separation of Machine Printed and HandWritten Text Lines”, Proc. of the 5th International Conference on Document Analysis and Recognition ( ICDAR) , vol. 336, pp. 37 − −45, 1999. [10] U. Pal and B. Choudhury, “Printed Devanagari Script OCR System”, KBCS-1997, vol. 336, pp. 37 − −45, 1997. [11] S. Duda, P. E. Hart, D. G. Stork. Pattern Classification, Second Edition. John Wiley, 2002. [12] Kandarpa Kumar Sarma, “Novel Feature Set for Neural Character Recognition ” , Proc. of the 5th International Symposium on Robotics and Automation, San Miguel Regla Hidalgo, Mexico , pp. 409 − −414, August, 2006.

600