Marathi And Sanskrit Word Identification By Using Genetic Algorithm

Transcription

1 Reviewed Paper Volume 2 Issue 12 August 2015 International Journal of Informative & Futuristic Research ISSN (Online): Marathi And Sanskrit Word Identification By Using Genetic Algorithm Paper ID IJIFR/ V2/ E12/ 049 Page No Subject Area Key Words Wiener filter, Curvelet transform, Genetic algorithm, Neural Network Computer Engineering Received On Accepted On Published On Priyanka Pradip Kulkarni 1 Sonal Patil 2 Ganesh Dhanokar 3 Research Scholar, Department Of Computer Engineering G.H.Raisoni Institute of Engineering and Management, Jalgaon-Maharashtra Assistant Professor Department Of Computer Engineering G.H.Raisoni Institute of Engineering and Management, Jalgaon-Maharashtra Assistant Professor Department Of Computer Engineering G.H.Raisoni Institute of Engineering and Management, Jalgaon-Maharashtra Abstract Character recognition plays an important role in many recent applications. Recently there is growing trend among worldwide researchers to recognize handwritten words of many languages and scripts. However, most of the current work in these areas is limited to English and a few oriental languages. The lack of efficient solutions for Indic scripts and languages such as Sanskrit has hampered information extraction from a large body of documents of cultural and historical importance. There are various feature extraction techniques such as Gradient feature extraction, Stroke method, Fourier descriptor and chain code histogram. A proper feature extraction technique can increase the recognition ratio. In this work, a curvelet transform and shape moment are investigated for Marathi and Sanskrit handwritten word identification system. Curvelet transform supports the edges and curve discontinuities. After the curvelet transform several group of curvelet coefficients are generated at different scale and angles. These coefficient are used with GLCM to compute features as contrast, correlation, homogeneity and energy of word image. Genetic algorithm is chosen for the optimization and finding the number of hidden nodes which uses as an input to neural network in classification step. After optimization, classification techniques have been used for training and testing purpose. Most of the researchers are used various classifiers for recognition of online handwritten word like Hidden Marko Model, K- Nearest Neighbour and Support Vector Machine. It is found that Neural Network gives better accuracy for recognition purpose. Copyright IJIFR

2 1. Introduction Handwritten character recognition (HCR) is a part of offline character recognition. Handwritten characters have infinite variety of style from one person to another person. Due to this wide range of variability; it is difficult to recognize by a machine. Although the research in Optical Character Recognition (OCR) has been going on for last few decades, the goal of this area is still out of reach. Most of the researchers have tried to solve the problems based on the image processing and pattern recognition techniques. OCR (Optical Character Recognition) is an active field of research in Pattern Recognition. OCR methodologies can be classified based on two criteria; data acquisition process which can be on-line or off-line and type of the text which is printed text or hand-written text[1] Devnagari is the most admired Indian script, used by more than 500 million people, which forms the basis for several Indian languages including Hindi, Sanskrit, Kashmiri, Marathi and so on. English character recognition is extensively studied by many researchers and various commercial systems are available for it. But in case of Indian languages, the research work is very limited due to the complex structure of the language. Normally, HCR can be divided into three steps namely pre-processing, feature extraction, segmentation and classification. Pre-processing stage is to produce a clean character image, it is can be used directly and efficiently by the feature extraction stage. Feature extraction stage is to remove redundancy from data. Segmentation stage is to increase the efficiency for next stage. A classification stage is to recognize characters or words.feature extraction in HCR is a very important field of image processing and objects recognition [2]. 1.1 Devanagari Character Set Devanagari is an Indian, syllabic alphabetic type of script that is used to write several languages like Sanskrit, Hindi, Marathi, Bhojpuri, Nepali, Konkani, Sindhi,Marwari, Pali, Maithli and many languages that are spoken in various parts of India. The word Devanagari is a combination of two words deva which means God and nagari which means urban establishment. Put together, these words mean Script of the Gods or Script of the urban establishmentthe basic set of symbols of Devanagari script consists of 13 vowels (or swar), 37 consonants (or vyanjan) as shown in Figure Word Identification Architecture Figure 1.1: (a). Devanagari Vowels (b). Devanagari Consonants 4589

3 Optical word identification involves many steps to completely recognize and produce machine encoded text. These phases are termed as: Pre-processing, Segmentation, Feature extraction, Classification. The architecture of these phases is shown in figure 1.2 and these phases are listed below with brief description [3]. Pre-processing Figure 1.2: Architecture of System The pre-processing phase normally includes many techniques applied for binarization, noise removal, skew detection, slant correction, normalization, contour making and skeletonization like processes to make character image easy to extract relevant features and efficient recognition. Feature Extraction Feature extraction is used to extract relevant features for recognition of characters based on these features. First features are computed and extracted and then most relevant features are selected to construct feature vector which is used eventually for recognition. The computation of features is based on structural, statistical, directional, moment, transformation like approaches. Classification Each pattern having feature vector is classified in predefined classes using classifiers. Classifiers are first trained by a training set of pattern samples to prepare a model which is later used to recognize the test samples. The training data should consist of wide varieties of samples to recognize all possible samples during testing. Some examples of generally practiced classifiers are- Support Vector Machine (SVM), K- Nearest Neighbour (K-NN), Artificial Neural Network (ANN) and Probabilistic Neural Network (PNN). 4590

4 2. Literature Survey Studies have been carried on recognition of different words. The most advanced and efficient OCR systems are designed for English, Chinese and Japanese like scripts and languages. In context to Indian languages and scripts, recently a significant research work is proposed. Most of Indian work on character recognition is dominated by Devanagari script, which is used in writing of Hindi, Marathi, and Nepali and Sanskrit languages. The work is also performed on compound word for recognition and classification purpose by extracting different features. Different classifiers also used in comparison to increase the accuracy rate of recognition. Namita Dwivedi have described recognition of Sanskrit word using Prewitt s operator for extracting the features from an image thinning process is applied in pre processing technique Thinning is an important pre-processing step in OCR. The purpose of thinning is to delete redundant information and at the same time retain the characteristic features of the image. Freeman Chain code is one of the representation techniques that is useful for image processing, shape analysis and pattern recognition fields is used with heuristic approach for feature extraction. Genetic algorithm is used for non-linear segmentation of multiple characters. This recognition model was built from SVM classifiers for higher level classification accuracy [1]. U. Pal, Wakabayashi and Kimura also presented comparative study of Devanagari handwritten character recognition using different features and classifiers [4]. They used four sets of features based on curvature and gradient information obtained from binary as well as gray scale images and compared results using 10 different classifiers as concluded the best results 74.74% and 75.17% for features extracted from binary and gray image respectively obtained with Mirror Image Learning (MIL) classifier. Sarbajit Pal et al.[5] have described projection based statistical approach for handwritten character recognition. They proposed four sided projections of characters and projections were smoothed by polygon approximation. Nikita Gaur and Dayashankar Singh et al.[6] have described gradient feature extraction approach for recognition of Sanskrit word they have used sobel operator for edge detection The Sobel operator is used in image processing, particularly with edge detection algorithms. Technically, it is a discrete differentiation operator, computing an approximation of the gradient of the image intensity function. The skeletonized and normalized binary pixels of English characters were used as the inputs of the MLP network. The results of structure analysis show that if the number of hidden nodes increases the number of epoches taken to recognize the handwritten character is also increases. A lot of efforts have been made to get higher accuracy as 94%. Brijmohan Singh, Ankush Mittal, M.A. Ansari, Debashis Ghosh et al.[7] have described a holistic system of offline handwritten Devanagari word recognition. In this paper, they proposed a Curvelet feature extractor with SVM and k-nn classifiers based scheme for the recognition of handwritten Devanagari words. Vedgupt Saraf, D.S. Rao et al.[8] have proposed a genetic algorithm scheme towards the recognition of off-line Devnagari handwritten characters. They tested their proposed system on different individuals samples and obtained 98.78% recognition accuracy. Prachi patil,saniya ansari et al.[9]have proposed Online Handwritten Devnagari Word Recognition using HMM based Technique. Feature extraction of input image is done by 4591

5 android technology. Using that features HMM recognizes the word. They tested proposed system on different word images and obtained 95.70% recognition accuracy. Gaurav kumar,pradeep kumar Bhatia et al.[10] have proposed Neural Network based Approach for Recognition of Text Images. They used Otsu method for binarization, after thet median filter is used for noise removal. Feature extraction takes place through Fourier transform. A multilayer feed forward neural network is created and trained through Back Propagation algorithm for classification achieves 93.58% accuracy. Sushama Shelke,Shaila Apte et al.[11] have proposed A Multistage Handwritten Marathi Compound Character Recognition Scheme using Neural Networks and Wavelet Features In this the recognition is carried out using multistage feature extraction and classification scheme. The average recognition rate is found to be 96.14% and 94.22% respectively for training and testing samples with wavelet approximation features and 98.68% and 96.23% respectively for training and testing samples with modified wavelet features. Sandhya Arora1. Debotosh Bhattacharjee et. al.[12] proposed Performance Comparison of SVM and ANN for Handwritten Devnagari Character Recognition. They extracted shadow features, chain code histogram features, view based features and longest run features. These features are then fed to Neural classifier and in support vector machine for classification. They achieved the accuracy rate as 99.62% through SVM and 99.50% using NN. M. N. Sandhya Arora, D. Bhattacharjee et. al.[13] proposed Recognition of non-compound handwritten devnagari characters using a combination of MLP and minimum edit distance they used two well-known and established pattern recognition techniques: one using neural networks and the other one using minimum edit distance and characters are represented using shadow feature and chain code histogram. The method is carried out on a database of 7154 samples. The overall recognition is found to be 90.74%. 3. Steps for Identification of Marathi and Sanskrit Word The basic steps for recognition of handwritten Devnagari Marathi and Sanskrit word recognition system are pre-processing, feature extraction and recognition. 3.1 Dataset The standard (benchmark) database for Indian script is neither available freely nor commercially, hence, Collect the samples of Devanagari handwritten words data from different professionals belonging to schools, colleges. The contributors to these data samples are of different educational backgrounds of metric, graduate, post graduate level qualification and different professions as student, teacher, and children. This database contains digital images. Consists of 624 samples for training purposes which are collected from 26 different writers of different profile and educational backgrounds. We have taken all these samples papers written in isolated manner. The writers wrote the word in different size using different pens to write. These words are having different character width. In some words there are distortions also introduced and amount of such distortion depends on quality of pen ink used to write words and speed of the writer to write the words. 3.2 Image pre-processing The raw data is subjected to a number of preliminary processing steps to make it usable in the descriptive stages of character analysis. Pre-processing aims to produce data that are easy for the character recognition systems to operate accurately. 4592

6 3.2.1 Smoothing Smoothing operations are used to blur the image and reduce the noise. Blurring is used in preprocessing steps such as removal of small details from an image. In binary images, smoothing operations are used to reduce the noise or to straighten the edges of the characters, for example, to all the small gaps or to remove the small bumps in the edges (contours) of the characters. Generally filters are used to filter unwanted things or object in a spatial domain or surface. In digital image processing, mostly the images are affected by various noises. The main objectives of the filters are to improve the quality of image by enhancing is to improve interoperability of the information present in the images for human visual. In this system wiener filter is used for smoothing Gaussian filter Noise in digital images can be described as random fluctuations in brightness and color. It will corrupt information in an image such that the intensity at each pixel is a combination of the true signal and the noise. Although many scanners and cameras have some built in noise r Reduction most digital images will display some degree of noise as a result of either Scanning,transmission or conversion Two of the most common types of noise are additive Gaussian and salt and pepper. Gaussian noise, or amplifier noise, will impair the image with a linear addition of white noise, meaning that it is independent of the image itself and usually evenly distributed over the frequency domain.as the name suggests, the intensity of the noise at each pixel follows a Gaussian normal distribution. It is most apparent for high frequencies and can in most cases be reduced by some kind of low pass filter. A Gaussian blur (also known as Gaussian smoothing) is the result of blurring an image by a Gaussian function to reduce image noise and reduce detail. Mathematically, applying a Gaussian blur to an image is the same as convolving the image with a Gaussian function Wiener Filter The inverse filtering is a restoration technique for deconvolution, i.e., when the image is blurred by a known low pass filter, it is possible to recover the image by inverse filtering or generalized inverse filtering. However, inverse filtering is very sensitive to additive noise. The approach of reducing degradation at a time allows us to develop a restoration algorithm for each type of degradation and simply combine them. The Wiener filtering executes an optimal trade-off between inverse filtering and noise smoothing. It removes the additive noise and inverts the blurring simultaneously. The Wiener filtering is optimal in terms of the mean square error. In other words, it minimizes the overall mean square error in the process of inverse filtering and noise smoothing. The Wiener filtering is a linear estimation of the original image [14]. The approach is based on a stochastic framework. The orthogonality principle implies that the Wiener filter in Fourier domain can be expressed as follows: Where Sxx(f1,f2),Sηη(f1,f2) are respectively power spectra of the original image and the additive noise, and H(f1,f2) is the blurring filter. It is easy to see that the Wiener filter has two separate part, an inverse filtering part and a noise smoothing part. It not only performs the deconvolution by inverse filtering (high pass filtering) but also removes the noise with a compression operation (low pass filtering). 4593

7 3.3 Feature Extraction Technique Any given image can be decomposed into several features. Feature extraction technique is accurately retrieve features of characters. The extracted features are organized in a database, which is the input for the recognition phase of the classifier. Feature extraction is a very important in recognition system because it is used by the classifier to classify the data Curvelet Transform A feature extraction scheme based on digital Curvelet transform has been used in[14]. In this work, the words from the sample images are extracted using conventional methods. A usual feature of handwritten text is the orientation of text written by the writer. Each sample is cropped to edges and resized to a standard width and height suitable for digital Curvelet transform. The digital Curvelet transform at a single scale is applied to each of the samples to obtain Curvelet coefficients as features. For large set of characters, as in Devanagari language, automatic curve matching is highly useful. Considering this, here the use of curvelet transforms which represents edges and singularities along curves more precisely with the needle-shaped basis elements. The elements own super directional sensitivity and smooth contours capturing efficiency. Since Curvelet are two dimensional waveforms that provide a new architecture for multiscale analysis, they can be used to distinguish similar appearing characters better. The Curvelet frame preserves the important properties, such as parabolic scaling, tightness and sparse representation for surface-like singularities of co-dimension one. Since many of the characters in a word not only consist of edge discontinuities but also of curve discontinuities. The most widely used Wavelet transform works well with edge discontinuities but a curve discontinuity affects all the Wavelet coefficients. On the other hand, the curve discontinuities in any character or word are well handled with Curvelet transform with very few numbers of coefficients. Hence, Curvelet-based feature are likely to work well for Devanagari character and word recognition. The curvelet transform includes four stages: i.) ii.) iii.) iv.) Sub-band decomposition: is to divide the image into resolution layers where each layer contains details of different frequencies. Smooth Partitioning: Each sub band is smoothly windowed into squares of an appropriate scale. Renormalization: Each resulting square is renormalized to unit square. Ridge let analysis: Each square is analyzed in the ortho-ridgelet system Curvelet are appropriate bases for representing images (or other functions) which are smooth apart from singularities along smooth curves, where the curves have bounded curvature, i.e. where objects in the image have a minimum length scale. The features extracted from input image by using curvelet transform as contrast, correlation, homogeneity and energy. a. Contrast:-Contrast is the difference in luminance or color that makes an object (or its representation in an image or display) distinguishable. In visual perception of the real world, contrast is determined by the difference in the color and brightness of the object and other objects within the same field of view. b. Correlation:- the process of establishing a relationship or connection between two or more things. c. Homogeneity:- the quality or state of being homogeneous. 4594

8 d. Energy:-It is a function that would capture the solution we desire and perform gradient descent to compute its lowest value, resulting in a solution for the image segmentation Shape Moment Feature In fact, the main problem in character recognition system is the large variation in shapes within a class of character. This variation depends on font styles, document noise, photometric effect, document skew and poor image quality. The large variation in shapes makes it difficult to determine the number of features that are convenient prior to model building. Various shape based and boundary based features are taken from individual character. The various Moment based features like total Kurtosis, Skewness, moment, percentile and quartile are calculated from the character images. 3.4 Optimization Genetic algorithms are a very good means of optimizations in such problems. They optimize the desired property by generating hybrid solutions from the presently existing solutions. These hybrid solutions are added to the solution pool and may be used to generate more hybrids. These solutions may be better than the solutions already generated. All this is done by the genetic operators, which are defined and applied over the problem [15]. Genetic algorithms are a good means of optimizations. Here proposed work use genetic algorithm to compute optimum number of hidden nodes which is the input to Multilayer Perceptron Neural Network. The system selects the optimum number of nodes when the mean square value becomes minimum at the same time the NN reaches the maximum number of epoch. Fitness Function: - In Genetic Algorithms, the fitness function is used to test the goodness of the solution. This function, when applied on any of the solution from the solution pool, tells the level of goodness. 3.5 Classification The decision making part of a recognition system is the classification stage and it uses the features extracted from the previous stage. There are various methods for classification. K Nearest Neighbor (KNN), Artificial Neural Network(ANN) and Support Vector Machine (SVM). The characteristics of the some classification methods that have been successfully applied to Offline Devnagari Marathi and Sanskrit word recognition and results of Neural Network classification is better than other classification methods, applied on Handwritten Devnagari word. 3.6 Artificial Neural Network The network needs to be trained first with some predefined standard character patterns to perform the recognition task. BPNN algorithm is used for this, which is considered as the unsupervised form of learning method where every neuron competes with each other in the basis of their activation value. The connection weights towards the winner neuron get adjusted during training process. Some random values are assigned initially to all the connection weights, during the training process these values are converged to some fixed values. The training process is similar to an unsupervised training method [16]. Artificial neural networks are one of the popular techniques used for classification due to their learning and generalization abilities. They have been traditionally used for character recognition application. Out of various architectures, multilayer perceptron (MLP) is widely used. The MLP is a fully connected network, where every neuron in a layer is connected to each and every neuron in the next layer by a weighted link through which the state of the neuron is transmitted. It consists of 4595

9 an input layer, a hidden layer and an output layer. A bias is similar to weight. It acts exactly as a weight on a connection from a unit whose activation is always one. Each neuron in the hidden layer includes a nonlinear activation function [11]. A multilayer perceptron (MLP) is a feed forward artificial neural network model that maps sets of input data onto a set of appropriate outputs. A MLP consists of multiple layers of nodes in a directed graph, with each layer fully connected to the next one. Except for the input nodes, each node is a neuron (or processing element) with a nonlinear activation function. MLP utilizes a supervised learning technique called back propagation for training the network. MLP is a modification of the standard linear perceptron and can distinguish data that are not linearly separable. When the network weights and biases are initialized, the network is ready for training. The multilayer feed forward network can be trained for function approximation (nonlinear regression) or pattern recognition. The training process requires a set of examples of proper network behaviornetwork inputs p and target outputs The process of training a neural network involves tuning the values of the weights and biases of the network to optimize network performance, as defined by the network performance function net.performfcn. The default performance function for feed forward networks is mean square error mse-the average squared error between the networks outputs a and the target outputs t. It is defined as follows: When training large networks, and when training pattern recognition networks, trainscg and trainrp are good choices. Their memory requirements are relatively small, and yet they are much faster than standard gradient descent algorithms. The optimisation technique used for training this architecture was the Scaled Conjugate Gradient (SCG) method. SCG method was used because it gave better results and has been found to solve the optimization problems encountered when training an MLP network more efficiently than the gradient descent and conjugate gradient methods [17]. 4. Result The experimental result is given below in detail. 4.1 Experimental Results In-order to train the classifiers, a set of training Marathi and Saskrit words are required and their respective features which are extracted through feature extraction step. There are two basic phases of pattern classification. They are training and testing phases. In the training phase, data is repeatedly presented to the classifier, while weights are updated to obtain a desired response. In testing phase, the trained system is applied to data that it has never seen to check the performance of the classification. Hence, we need to design the classifier by partitioning the total data set into training and testing data set. From the total data set, In neural network before training the data the whole data needs to partition into training data, validation testing data and testing data. Each time when the neural network train the whole data divide into these three types. Each time a neural network is trained can result in different solution due to different initial weight and bias value and different division of data into training, validation and test sets. That s why each time it gives different output to the same input. The table given in the 4.1 below shows the percentage accuracy for sample five word images from Marathi word dataset and 4.2 shows the percentage accuracy for 4596

10 sample five word image from Sanskrit word dataset. The accuracy for both Marathi and Sanskrit words are different each time when we train the neural network. Table 4.1: Accuracy Result Table for Marathi Word Sample Table 4.2: Accuracy Result Table for Sanskrit Word Sample 5. Conclusion In research work standard datasets for Marathi and Sanskrit word is not publically available, since first of all prepared database of 624 samples for Marathi and Sanskrit word by considering all possible constraints such as variations in writing styles, similar word sample consisting with some noise etc. Words are written from different people of different age. The technique discussed above is curvelet transform and shape moment involves the extraction of the shape feature. Feature 4597

11 extraction for compound word is difficult. But curvelet transform supports the edges and curve discontinuities. Hence, Curvelet Transform proves to be useful in Marathi and Sanskrit word recognition system. This work present neural network based classification system for the recognition of offline handwritten Marathi and Sanskrit words. Among all the methods it is found that NN gives better accuracy for recognition purpose. However, the accuracy of proposed scheme can be enhanced by increasing the number of training samples or applying the proposed scheme at different resolution scheme. This recognition system can be extended for the recognition of other words, sentence and documents. 6. References [1] N. Dwivedi and N. arya, Sanskrit Word Recognition using Prewitts Operator and Support Vector Classification, IEEE ICECCN, [2] B.Indira and R. Saqib, Devanagari Character Recognition: A Short Review, International Journal of Computer Applications, vol. 4, December [3] S. Kubatur and M. Sid-Ahmed, A Neural Network Approach to Online Devanagari Handwritten Character Recognition, IEEE, pp , December [4] U. Pal and F. Kimura, Comparative Study of Devanagari Handwritten Character Recognition using Different Feature and Classifiers, 10th International Conference on Document Analysis and Recognition, [5] S. Pal and J. Mitra, A Projection Based Statistical Approach for Handwritten Character Recognition, Proceedings of International Conference on Computational Intelligence and Multimedia Applications, vol. 2, pp , [6] N. Gaur and D. Singh, Sanskrit Word Recognition using Gradient Feature Extraction, VSRD-IJCSIT, vol. 2, pp , [7] B. Singh and A. Mittal, Handwritten Devanagari Word Recognition: A Curvelet Transform Based Approach, International Journal on Computer Science and Engineering (IJCSE), vol. 3, April [8] V. Saraf and D. Rao, Handwritten Devanagari Character Recognition using Gradient Features, International Journal of Soft Computing and Engineering (IJSCE), vol. 2, April [9] P. Patil and S. Ansari, Online Handwritten Devnagari Word Recognition using Hmm Based Technique, International Journal of Computer Applications, vol. 17, June [10]G. kumar and P. kumar Bhatia, Neural Network Based Approach for Recognition of Text Images, International Journal of Computer Applications, vol. 14, January [11] S. Shelke and S. Apte, A Multistage Handwritten Marathi Compound Character Recognition Scheme using Neural Networks and Wavelet Features, International Journal of Signal Processing, Image Processing and Pattern Recognition, vol. 1, March [12] S. Arora and D. Bhattacharjee, Performance Comparison of SVM and ANN for Handwritten Devnagari Character Recognition, IJCSI International Journal of Computer Science, vol. 7, May [13] S. Arora and D. Bhattacharjee, Recognition of Non-compound Handwritten Devnagari Characters using A Combination of MLP and Minimum Edit Distance, IJCSS. [14] S. Pannirselvam and S.Ponmani, Preprocessing of Handwritten Documents using Various Filters A Survey, International Journal of Advanced Research in Computer Science and Software Engineering, vol. 3, July [15] R. Kala and H. Vazirani, Offline Handwriting Recognition using Genetic Algorithm, IJCSI International Journal of Computer Science Issues, vol. 1, March [16] R. Dineshkumar and J. Suganthi, Sanskrit Character Recognition System using Neural Network, Indian Journal of Science and Technology, vol. 8, pp , January [17] M. Abdella and T. Marwala, The Use of Genetic Algorithms and Neural Networks to Approximate Missing Data in Database, Computing and Informatics, vol. 24, pp ,