Вы находитесь на странице: 1из 20

Neural Network based Feature Extraction for Assamese Character and Numeral Recognition

Kandarpa Kumar Sarma Department of Electronics Science Gauhati University Guwahati-781014, India kandarpaks@gauhati.ac.in

ABSTRACT Feature extraction is a vital ingredient in a neural network based pattern recognition problem. In problems like character and numeral recognition using neural network the feature extraction process should capture not only the relevant information about an input pattern but also ensure that only the effective samples are retained and used for subsequent operations. The work here describes the use of a neural feature extractor for generating a feature set for use in a neural network based character and numeral recognition system in Assamese- an important language in the North Eastern part of the country. The results thus obtained are compared with the ones earlier obtained from a similar application and a PCA-based feature extraction method. The neural feature extractor based method proves to be superior. Keywords: MLP, Feature, Correlation Recognition, Assamese.

1 Introduction
The feature vector of an input pattern like a character represents a unique set of data providing relevant information regarding shape, size, morphology etc. In a neural network based optical character recognition (OCR) system the feature vector should be formulated in such a way that it maximizes performance. In an OCR system designed using a neural network like the multi layered perceptron (MLP) the training stage suffers if the feature set is not properly formulated. One of the requirements of this nature is lower correlatedness among the feature samples. Lesser the correlation between the training samples better is the training and the subsequent recognition performance. A robust feature set directly aids the convergence of the mean square error (MSE) of a neural network during training. A feature set earlier formulated as reported in (Sarma Bora and Mahnata, 2005), (Sarma, 2006) and (Sarma, 2007) , proved to be robust enough for applications in character and numeral recognition. But still there remained scope for experimenting with a neural network congured as a feature extractor. In such a role the neural network can generate the best set of features ((Haykin, 2003), (Bishop, 2003)) required to perform training of another neural network which later can be used to perform recognition. A neural network applied for capturing the best set of features acts as

a feature extractor. The process of feature extraction is somewhat different than that of feature selection as noted by Jain, Duin and Mao in one of their reviews (Jain, Duin and Mao, 2000), (O. D. Trier and Tax, 1996). They dened feature selection as the process to sort out the best subset of the input features while the algorithms that create new features based on transformations or combinations of the original feature sets constitute feature extraction. Moreover the process of mapping original features (measurements) into fewer, more effective features is feature extraction (Lerner, Guterman, Aladjem and Dinstein, 1999). A neural feature extractor can be supervised or unsupervised. In this work a supervised neural network namely an MLP is used as a feature extractor. The feature extractor MLP extracts the feature set from the input characters which is next used to train another MLP used as the classier. The input set consists of Assamese characters and handwritten Anglo-Assamese numerals. The results obtained are compared to that obtained by using a feature set as re ported in (Sarma et al., 2005), (Sarma, 2006) and (Sarma, 2007) and another MLP based recognition using PCA based feature extraction. The inputs are applied to the MLP feature extractor which generates a feature vector for each of the input classes which are next used to train the classier- another MLP. Three primary reasons are behind the use of the neural feature extractor in this work. First, is to save time in the feature extraction phase as neural networks are fast discrimination and decision making tools. The second reason is to generate a robust feature set which can be used for training the MLP classier in the next stage. Finally, a neural feature extractor proves to be invariant to presence of noise in the input pattern and generates the desired results if the training is carried out properly. Marginal time saving and performance improvements have been recorded as compared to earlier attempts made without using the neural feature extractor. The scheme ( Figure 2) has been tested for the best mean square error (MSE) convergence rate and classication performance using multiple MLP congurations, learning rates and momentum constants. The idea is to determine the most suitable combinations of MLP conguration for use as feature extractor and recognizer. Considerable work has been done in Optical Character Recognition (OCR) of Indian lan guages. A detailed account of such work is available in (Sarma et al., 2005). Some of the previous work done in the eld of numeral recognition in Indian languages include works done in Bengali (Pal and Chaudhuri, 2000) and (Bhattacharya, Das, Datta, Pauri and Chaudhuri., 2000), Devnagari and English handwritten numeral recognition (Lehal and Bhatt, 2000) etc. This paper is organized into the following sections. Section 2 provides the mathematical basis of MLP based classication and training using (error) back propagation. In Section 3 characteristics of Assamese Numerals and related recognition challenges are discussed. Section 4 describes the algorithmic steps followed in the work. Details of the feature extraction process are explained in Section 5. Description of the formulation of the neural feature extractor is given in Section 5.1. Conguration of the MLPs and training are explained in Section 5.4. Performance details and results are available in Section 5.5. A conclusion is included in Section 6.

Classication using neural networks

Accuracy of a recognition process depends to a large extent on a higher value of proper classication. Neural networks has been one of the preferred classiers for their ability to provide optimal solution to a wide class of arbitrary classication problems(Haykin, 2003), (Duda, Hart and Stork, 2002), (Bishop, 2003). Multi Layered Perceptron (MLP)s trained with (error) Back Propagation (BP) in particular has found wide spread acceptance in several character recognition works. The reason is that the MLPs implement linear discriminants but in a space where inputs are mapped nonlinearly. The strength of this classication is derived from the fact that MLPs admit fairly simple algorithms where the form of nonlinearity can be learned from training data. The classication process involves dening a decision rule so as to provide boundaries to separate one class of data from another. It means placing the j th input of the input vector x(j) in the proper class among M outputs of a classier. Classication by a neural network involves training it so as to provide an optimum decision rule for classifying all the outputs of the network. The training should minimize the risk functional (Haykin, 2003), (Bishop, 2003): 1 (dj F (xj ))2 2N where dj is the desired output pattern for the prototype vector xj , R= sented to the network in training. The decision rule therefore can be given by the output of the network: ykj = Fk (xj ) for the jth input vector xj . The MLPs provide global approximations to non-linear mapping from the input to the output layers. As a result, MLPs are able to generalize in the regions of the input space where little or no data are available (Haykin, 2003). The size of the network is an important consideration from both the performance and computational points of view. (2.2) (2.1)

((.)) is the Euclidean norm of the enclosed vector and N is the total number of samples pre-

2.1 Multi Layered Perceptron Based Learning


The fundamental unit of the neural networks is the McCulloch-Pitts Neuron (1943). The MLP is the product of several researchers: Frank Rosenblatt (1958), H. D. Block (1962) and M. L. Minsky with S. A. Papart (1988). Backpropagation, the training algorithm , was discovered independently by several researchers (Rumelhart et. al.(1986) and also McClelland and Rumelhart (1988)). A simple perceptron is a single McCulloch-Pitts neuron trained by the perceptron algorithm is given as: Ox = g(([w].[x] + b) (2.3)

where [x] is the input vector,[w] is the associated weight vector,b is a bias value and g(x) is the activation function. Such a setup, namely the perceptron will be able to classify only linearly

Figure 1: MLP conguration to classify non-linearly distributed classes separable data. A MLP, in contrast, consists of several layers of neurons. The equation for output in a MLP with one hidden layer is given as:
N

Ox =
i=1

i g[w]i .[x] + bi

(2.4)

where i is the weight value between the ith hidden neuron, [w] is the vector of weights between the input and the hidden layer, [x] is the vector of inputs, [b] is the input bias of the hidden neuron layer and g(x) is the activation function. Such a set-up maybe depicted as in Figure 1. The process of adjusting the weights and biases of a perceptron or MLP is known as training. The perceptron algorithm (for training simple perceptrons consists of comparing the output of the perceptron with an associated target value. The most common training algorithm for MLPs is error backpropagation. This algorithm entails a back propagation of the error correction through each neuron in the network.

2.2 Application of Error Back Propagation for MLP training


The MLP is trained using (error)Back Propagation(BP) depending upon which the connecting weights between the layers are updated. This adaptive updating of the MLP is continued till the performance goal is met.Training the MLP is done in two broad passes -one a forward pass and the other a backward calculation with error determination and connecting weight updating in between. Batch training method is adopted as it accelerates the speed of training and the rate of convergence of the MSE to the desired value (Haykin, 2003) (Bishop, 2003). The steps are as below: Initialization: Initialize weight matrix W with random values between [-1,1] if a tansigmoid function is used as an activation function and between [0, 1] if log-sigmoid func-

tion is used as activation function. W is a matrix of CxP where P is the length of the feature vector used for each of the C classes. Presentation of training samples: Input is pm = [pm1 , pm2 .....pmL ]. The desired output is dm =[dm1 , dm2 ......dmL ]. Compute the values of the hidden nodes as:
L

neth mj

=
i=1

h wji pmi + h j

(2.5)

Calculate the output from the hidden layer as


h oh = fj (neth ) mj mj

(2.6)

where f(x)= e1 x
x

or f(x)= ex ex e +e depending upon the choice of the activation function. Calculate the values of the output node as:
o oo = fk (neto ) mj mk

(2.7)

Forward Computation: Compute the errors: ejn = djn ojn Calculate the mean square error(MSE) as : M SE = Error terms for the output layer is:
o mk = oo (1 oo )emn mk mk M j=1 L 2 n=1 ejn

(2.8)

2M

(2.9)

(2.10)

Error terms for the hidden layer:


h mk = oh (1 oh ) mk mk j o o mj wjk

(2.11)

Weight Update: Between the output and hidden layers


o o o wkj (t + 1) = wkj (t) + mk omj

(2.12)

where is the learning rate(01). For faster convergence a momentum term()maybe added as:
o o o o wkj (t + 1) = wkj (t) + mk omj + (wkj (t + 1) wkj )

(2.13)

Figure 2: Block diagram of the work showing the presence of the neural feature extractor Between the hidden layer and input layer:
h h h wji (t + 1) = wji (t) + mj pi

(2.14)

A momentum term maybe added as:


h h h o wji (t + 1) = wji (t) + mj pi + (wji (t + 1) wji

(2.15)

One cycle through the complete training set forms one epoch. The above is repeated till MSE meets the performance criteria. While repeating the above the number of epoch elapsed is counted. The methods used for MLP training includes: Gradient Descent( GDBP ) Gradient Descent with Momentum BP( GDMBP ). Gradient Descent with Adaptive Learning Rate BP( GDALRBP ). Here, while MSE approaches the convergence goal, training of the MLPs suffer if: Corr(pm,i (j), pm,i (j + 1)) = high and Corr(pm,i (j), pm,i+1 (j))= high. This is due to the requirements of the (error) back propagation algorithm as suggested by equations 2.10 and 2.11.

Assamese Characters, 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 (Chaudhuri and Pal, 1998) of the characters and numerals. Some characteristic features of Assamese characters are available in (Sarma et al., 2005). A sample set of Assamese characters used for the training of the MLPs for performing the combined character and numeral recognition is shown in Figure 3. 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 4).

Figure 3: Assamese character set used for training MLP classier

Figure 4: Assamese numerals used for training MLP classier

Figure 5: Handwritten Assamese numeral samples used for training

Figure 6: Assamese numerals and equivalent English forms Handwritten numerals can contain huge amount of variations due to differences in writing styles, orientation etc (Figure 5). Writer variation results in modication 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. Handwritten numerals can contain huge amount of variations due to differences in writing styles, orientation etc (Figure 5). Writer variation results in modication 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.

Figure 7: Handwritten Assamese and English numerals

4 Algorithmic steps of the present work


The entire system maybe described by the following algorithmic steps: Training: 1. Generation of feature vectors of all possible classes. Total number of classes amount to 109. In this case 11 vowels, 41 consonants, 19 compound characters, 18 special characters ( Figure 3 ) in Assamese and 10 numerals each of Assamese and English have been used ( Figures 7 and 6). First, a feature set of all of these 109 classes have been generated by following the method as reported in (Sarma et al., 2005), (Sarma, 2006) and (Sarma, 2007). That feature set is next used to train two MLP congurations for application as feature extractor. The feature set generated by the MLP feature extractor is next used to train the MLP classier. 2. Designing and training the MLPs in three congurations (with error Back Propagation (BP)) with four learning rates and two momentum constants. The three different congurations are taken to ascertain the best conguration with respect to performance. 3. Performing classication for all available classes of characters and numerals used for training the MLPs in all the congurations. The MLP with the best MSE convergence and classication performance is used to perform recognition. 4. Performing classication for all available classes of characters and numerals used for training the neural network with noise mixed data. The convergence of the MSE and classication performance are noted. 5. Determining the best MSE convergence and classication performance with respect to feature type and MLP conguration. Testing: 1. After pre-processing, segmentation is performed. At rst, segmentation is done to obtain lines from page, next words from lines and nally letters and characters from words. A similar segmentation is carried out for the numerals given to the system as text input.

2. Then feature vectors are generated for the testing set by feeding the test inputs directly to the selected MLP feature extractor. The test set taken includes multiple font and size variations with noise free and noise mixed forms. 3. The feature vectors thus generated for each of the test patterns are 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.

Feature Extraction

The detailed description of the feature extraction process with respect to the formation of the hybrid feature set for use in combined character and numeral recognition is avail able in (Sarma et al., 2005) and its modication for exclusive use in numeral recognition is available in (Sarma, 2006) and (Sarma, 2007). The original hybrid feature set is of length 223 (Sarma et al., 2005) which makes it suitable enough for handling upto 111 classes of input by following the norm of assigning two features samples per input (Pal and Choudhury, 1999a), (Pal and Choudhury, 1999b). In this work the total number of classes amounts to 109. The features used to perform the training of the MLP based feature extraction is derived following the descriptions available in (Sarma et al., 2005), (Sarma, 2006) and (Sarma, 2007). The feature set thus derived is used rstly for training the MLP classier and subsequently for recognition of a combined inputs of Assamese characters and Anglo-Assamese numerals. The results of the experiments carried out is provided below in Section. 5.5.

5.1

Neural Feature Extractor Design

The process of formulation of the proper neural feature extractor is based on two MLPcongurations one with two hidden layers and the other with three hidden layers. These two congurations are designed to have an output layer of size 223 for 109 classes for the combined character and numeral recognizer. The conguration of the two MLPs maybe shown as in Table 1. The input layer length is determined by the size of the patterns applied to the MLP feature extractors. Three sizes of patterns have been considered as inputs to the extractor congurations. These are 8x8, 16x16 and 32x32. The results generated using the two MLP congurations trained for 7000 epochs with Gradient descent with momentum back propagation and tested with a sample size of 100 for each of the classes are shown in Tables 2 and 3. The Table 2 shows the results of feature extraction by a 2-hidden layered MLP receiving three variations of input of sizes 8x8, 16x16 and 32x32 and generating a feature set of length 223 for 109 classes.

Table 1: Conguration of the MLPs for neural feature extractor and classier Cases Hidden Layers Neurons/hidden layer P-input layer length 1 2 Two Three 1.5*P, 0.89*P 1.5*P, 0.89*P 0.6 *P Table 2: Comparison measure of xing the input pattern size of MLP feature extractor with two hidden layers for combined character and numeral recognition Cases Pattern MSE attained Classication Performance Size 1 2 3 8x8 16 x 16 32 x 32 3 x 104 2.1 x 1.9 x 104 104 in % 91 93.2 96.5 Time taken in sec.s per class 23 31 33

A similar set of results for a three hidden layered MLP feature extractor is shown by Table 3. Tables 2 and 3 show that the two hidden layered MLP form has better MSE convergence, classication rate and time- performance during the feature extraction phase. Also, the 32x32 input pattern size produces the best MSE convergence, classication performance and time consumed per class in training during the formulation of the neural feature extractor. Hence, the pattern size applied to the feature extractor is xed at 32x32 and all test inputs are normalized to that dimension. The work described in (Sarma, 2007) shows that a feature vector of size 50 derived out of the original 223 length feature is suitable for numeral recognition. A trimmed and modied feature set of 20 Anglo-Assamese numerals have been derived from the original 223 length feature for training another set of neural feature extractors for use exclusively for numeral recognition. A table showing the effect of feature length on MLP training for handwritten numeral recognition is reproduced here (Table 4) which has already been reported in (Sarma, 2007). Multiple MLP feature extractors have also been congured for generating feature sets for

Table 3: Comparison measure of xing the input pattern size of MLP feature extractor with three hidden layers for combined character and numeral recognition Cases Pattern MSE attained Classication Performance Time taken in sec.s Size 1 2 3 8x8 16 x 16 32 x 32 4.1 x 3.8 x 2.3 x 104 104 104 in % 89.4 90.2 93.5 per class 32 38 45

Table 4: Effect of feature vector length on MSE Convergence and Recognition rate using a 2-hidden layered MLP. Case Feature Length Average MSE Convergence Average Class wise Recognition in % 1 2 3 4 5 6 7 8 9 10 223 200 170 140 110 80 50 40 30 20 2x103 1.22x103 1.72x104 2.12x105 3.02x105 2.12x105 0.12x105 1.7x104 1.61x104 1.43x104 85.2 83.6 77.8 78.6 79.3 78.6 95.6 91.6 90.6 91.6 Av. intersample varia -nces of feature sets in % 1.5 3 6 9 12 15 21 24 27 30

Table 5: Comparison measure of xing the input pattern size of MLP feature extractor with two hidden layers for combined character and numeral recognition Cases Pattern MSE attained Classication Performance Size 1 2 3 8x8 16 x 16 32 x 32 4.8 x 6.7 x 105 105 in % 93 94.6 97.5 Time taken in sec.s per class 13 22 24

3.1 x 105

recognition of handwritten numerals exclusively (Tables 5 and 6). The 2-hidden layered MLP feature extractor receiving a 32 x 32 input generates the better MSE convergence and classication rate (Tables 5 ), hence will be utilized as the standard feature extractor for numeral recogntion.

5.2

PCA-feature generation

Principal Component Analysis (PCA) involves a mathematical procedure that transforms a number of correlated variables into a (smaller) number of uncorrelated variables called principal components. The rst principal component accounts for as much of the variability in the data as possible, and each succeeding component accounts for as much of the remaining variability as possible (Bockw, 1998). PCA of an input therefore represents an important description of a pattern and can be used as a feature set. PCA also contributes towards generating an optimized set of samples capturing relevant information of an input. PCA of characters and numerals are therefore used to train MLP classiers to

Table 6: Comparison measure of xing the input pattern size of MLP feature extractor with three hidden layers for combined character and numeral recognition Cases Pattern MSE attained Classication Performance Time taken in sec.s Size 1 2 3 8x8 16 x 16 32 x 32 5.3 x 7.2 x 105 105 in % 91.8 92.5 96.4 per class 15 24 27

4.8 x 105

generate certain results for serving as a comparison measure to the present work.

5.3

Combined Character and Numeral Recognition using MLPs

In this work, MLPs in two congurations are trained for classication with a set of features generated by a trained MLP-extractor. The inputs consists of 89 Assamese characters and 20 Anglo-Assamese handwritten numerals. The feature vectors are generated from the scanned inputs of the patterns applied to the MLP extractors for the combined character-numeral recognizer and exclusive numeral recognizer. The specic MLP classiers are therefore trained to handle upto 109 types of combined inputs and another 20 numerals separately. The training data set is derived during the feature extraction phase from a set which consists of 1000 samples per input as shown in Figures 3 and 7. The entire work has two part: training and testing. During training, feature vectors of all the classes of characters and numerals are generated by the feature extractors. These are next used to train the MLP classiers. Two different MLP are formed and trained with four learning rates and two momentum constants to ascertain the best conguration with respect to MSE convergence and classication performance. The network with the best MSE convergence and classication performance is used to perform recognition. During testing, rst, 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 applied to the neural extractors taking several samples so that the content includes multiple 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.

5.4

Conguration of the MLPs and Training

One hidden layer of an MLP is sufcient to approximate the mapping of any continuous function (Bishop, 2003). However, with many class problems like the present one, multiple hidden layered MLPs provide better MSE convergence, classication 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 classication (Bishop, 2003). Several congurations 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; but the size of the input layer depends upon the length of the feature vector and the length of the output layer reects the number of classes. The conguration of the MLPs used for the combined Character and Numeral Recognition System maybe shown in a tabular form as in Table 1 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 (Haykin, 2003). Similarly, for the exclusive numeral recognition system, for twenty classes, a single hidden layered-MLP is sufcient. But handwritten numeral recognition is a task with considerable complexity which requires multiple hidden layered-MLP to perform the classication and recognition. Multiple hidden layered-MLPs learn the writer-induced variations of the inputs and deal with related variations.

5.5 Performance of the recognition system trained using Feature Set generated by the neural feature extractor :
Classication 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 classication is 186 then the successful classication rate is taken to be 93%. A summary of the results using the neural feature extractor and the improvements achieved compared to the case without using the neural feature extractor and PCA based features is shown in Table 7. The MLPfeature extractor provides improvement of 1.5% in recognizing characters compared to the case without using the neural extractor. This measure is 4.6% compared to the case when PCA-features are used to train the MLP-classiers. In other cases also similar improvements in performance is achieved. The highest improvement of 4.9% is achieved during combined character and numeral recognition compared to PCA-features. This improvement is due to the robustness of the feature set generated by the neural feature extractor. The original hybrid features showed higher levels of uncorrelatedness in the training data. The neural feature extraction contributes further towards the improvement of that robustness and uncorrelatedness. 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 xed during training. All measures must be taken to make the uncorrelatedness

Table 7: Average Recognition results showing improvements achieved using MLP-feature extractor compared to the case without using the neural feature extractor and a PCA-based feature extraction method. Cases Using MLPFeature Extractor in (%) (1) Characters Assamese Numerals English Numerals AngloAssamese Numerals Combined Character and numeral input as low as possible. The features thus generated appears to be robust as an improvement to the original hybrid feature set reported in (Sarma et al., 2005), (Sarma, 2006) and (Sarma, 2007) this time in dealing with both characters and numerals together. The 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 around 96% as reported in (Sarma, 2007) which have not been included here. The rate of correct recognition is around 97% (Figure 8) in case of Assamese handwritten numerals. The percentage of miss hits shows a decrease compared to results obtained by using the original feature of length 223 which reects the spinoffs of the data pruning. A similar type of recognition rate (around 97% ) is achieved when only English numerals are used as input to the trained MLP (Figure 9). Another set of results derived by using mixed numeral input to the selected MLP point to some problems associated with the system (Figure 10). The system fails to make a distinction between the Assamese 4 and the English 8. The system recognizes Assamese 4 as English 8 and reverse. It is a drawback and needs to be addressed. There are similar high percentage of miss hits recorded when Assamese 0 is used which is recognized as English 0 and vice versa. This however will have no adverse effect on the performance of the system. 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 nd out the robustness of the MLP. Noise, with variations of 10dB, 30dB, 50dB, 70dB and 90dB, is mixed with 95.8 94.6 91.3 1.3 4.9 97.7 96.3 94.3 1.5 3.6 97.9 96.1 94.5 1.9 3.6 96.3 97.8 Without using MLP- Feature Extractor in (%) (2) 94.8 96.2 Using PCA-based Features in (%) 92.1 93.7 Improvement in (%) compared to 1 1.6 1.7 Improvement in (%) compared to 2 4.6 4.4

Figure 8: Recognition results of Assamese numerals

Figure 9: Recognition results of English numerals

Figure 10: Recognition results of mixed numeral inputs the input text data from where individual characters are segmented out before passing the characters to the feature extraction process. Table 8 shows the effect of noise on the classication and recognition process obtained with the variation of MSE, Classication and Recognition Rates for varying signal to noise ratio (SNR) without using the neural feature extractor. The advantage of using the neural feature extractor can be seen in case of numeral recognition from the Table 9 which clearly shows the improvements achieved. The classication rate is an average value obtained with two different MLP congurations 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 neural feature extractor is obvious. The training of the feature extractor however is crucial as it effects the subsequent recognition performance of the system. Table 10 throws some light in this direction. The results shown here been obtained by following trial and error method. Therefore, the training conditions of the neural feature extractor needs to be carefully xed before proceeding to the next stages. The training of the feature extractor must be done by keeping an eye on the classication and recognition performance of the system. The choice of 7000 epochs for training of the MLP-feature extractor and the classier is distinctly established by the Table 10.

Conclusion

A robust feature set is vital in all forms of neural recognition problems. It is true with character and numeral recognition systems. The robustness of the feature set to a large extent determines the success of the recognition process. As neural networks are adap-

Table 8: MSE, Classication and Recognition Rate variation for varying SNR values for a 2-hidden layered MLP with momentum constant, = 0.6 used for handwritten numeral recognition without using MLP-feature extractor SNR MSE Classication in dB 10 30 50 70 90 after 5000 epochs 0.00051 0.00042 0.00043 0.00040 0.00036 81.3 86.6 88.1 93.2 96.6 86.4 88.3 89.8 95.1 97.2 Rate (%) Rate (%) Recognition-

Table 9: MSE, Classication and Recognition Rate variation for varying SNR values for a 2-hidden layered MLP with momentum constant, = 0.6 used for handwritten numeral recognition using MLP-feature extractor SNR MSE in dB 10 30 50 70 90 after 5000 epochs 0.0004 0.00032 0.00029 0.00028 0.00022 83.7 89.8 90.3 94.1 97.2 87.1 89.2 90.2 96.2 97.3 Classication Rate (%) Rate (%) Recognition-

Table 10: Effect of variation of training conditions of neural feature extractor on average recognition performance Case Epochs Average MSE 1 2 3 4 5 6 7 8 9 10 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 2.2 x 3.3 x 1.7 x 1.2 x 0.8 x 0.3 x 0.33 x 0.385 x 102 103 103 105 105 105 105 105 Average RecognitionRate (%) 82.4 87.3 89.2 90.5 91.7 94.2 95.3 94.8 94.2 93.7

5.1 x 104

0.38 x 105

tive learning tools, these can work effectively as feature extractors. In such a role a neural network generates the optimized set of features for improving recognition performance of a neural network based recognition system. The work here showed the effectiveness of use of a neural feature extractor trained with a hybrid feature set to improve performance of a character and numeral recognition system in Assamese. The success rate achieved is healthy. However, inter-language mishits have affected the system performance for which hybrid classication mechanisms needs to be formulated to increase the success rate of the work.

References
Bhattacharya, U., Das, T. K., Datta, A., Pauri, S. K. and Chaudhuri., B. B. 2000. A Hybrid Scheme for Handprinted Numeral Recognition based on a Self-Organizing Network and MLP Classiers, IJPRAI 16(7): 845864. Bishop, C. M. 2003. Neural Networks for Pattern Recognition, First Indian Edition . Bockw, R. K. 1998. Principal Component Analysis, http://rkb.home.cern.ch/rkb/AN16pp/node215.html (April). Chaudhuri, B. B. and Pal, U. 1998. A Complete Printed Bangla OCR System, Pattern Recognition 31(5): 531549. Duda, S., Hart, P. E. and Stork, D. G. 2002. Pattern Classication, Second Edition . Haykin, S. 2003. Neural Networks A Comprehensive Foundation, Second Edition . Jain, A., Duin, R. and Mao, J. 2000. Statistical Pattern Recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence 22(Jan): 435. Lehal, G. S. and Bhatt, N. 2000. A Recognition System for Devnagri and English handwritten numerals, T. Tan, Y. Shi, and W. Gao (Eds.): LNCS 1948 pp. 442449. Lerner, B., Guterman, H., Aladjem, M. and Dinstein, I. 1999. A Comparative Study of Neural Network based Feature Extraction Paradigms, Pattern Recognition Letters 20(1): 714. O. D. Trier, A. J. and Tax, T. 1996. Feature Extraction Methods for Character Recognition, Pattern Recognition 29(4): 641 662. Pal, U. and Chaudhuri, B. B. 2000. Automatic Recognition of Unconstrained Off-line Bangla Handwritten Numerals, T. Tan, Y. Shi, and W. Gao (Eds.): LNCS 1948 pp. 371378. Pal, U. and Choudhury, B. 1999a. Automatic Separation of Machine Printed and HandWritten Text Lines, Proc. of the 5th International Conference on Document Analysis and Recognition ( ICDAR) 336: 3745. Pal, U. and Choudhury, B. 1999b. Printed Devanagari Script OCR System, KBCS - 1997 336: 3745. Sarma, K. K. 2006. Novel Feature Set for Neural Character Recognition, 5th International Symposium on Robotics and Automation, San Miguel Regla Hidalgo, Mexico.

Sarma, K. K. 2007. MLP-based Assamese Character and Numeral Recognition using an Innovative Hybrid Feature Set, Proceedings of 3nd Indian International Conference on Articial Intelligence (IICAI-2007) pp. 585600. Sarma K. K., Bora, P. K. and Mahnata, C. 2005. Innovative Feature Set for Multi Lay, ered Perceptron (MLP) based Assamese Character Recognition, Proceedings of 2nd Indian International Conference on Articial Intelligence (IICAI-2005) pp. 473491.

Вам также может понравиться