Offline arabic handwriting recognition and technology

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Chapter 1


Chapter 2 Literature Review

2.1 Introduction

Despite the widespread use of electronic files, paper documents are still playing an important role in our daily lives. Technological development has allowed us to convert the paper documents to electronic documents in the form of images using some devices such as scanner. Because of that many of the research conducted on the text recognition in the last two eras.

Handwriting recognition is one of the difficult topics because of its variations and inconsistencies. It is divided onto two basic types' offline and online recognition. Offline Arabic handwriting recognition is vital issue that currently being

The sixties and seventies were the period in which the works on the recognition of handwriting began. The performance of the built systems at that time was very powerless. This was the reason for the lack of research on the same subject in the eighties(Jay Ashree R.Prasad, 2010).

2.2 Trends Of Handwriting Recognition

In 1950's all the trends were directed to the field of segmentation where the ability to segment the original image was considered as indicator of the quality measurement (Jay Ashree R.Prasad, 2010). In 1960's integrated segmentation and interpretation systems were the main methodologies of this period because it witnessed the appearance of some high quality equipments which was used specially in image acquisition (Jay Ashree R.Prasad, 2010).

2.2.1 1970's

evaluation methods are the main procedures of these techniques. The search method aims to generate candidate sets. The evaluation method used to evaluate the selected feature depending on a performance index. Applying this technique on eight-class characters in addition of 19.000 typed characters as database led to satisfactory results.

Kozlay (1971) developed a new feature extracting technique which characterized by using small number of equipments. It just needs a scanner, minicomputer, tape and hardware correlator. This technique based on collecting the maximum number of linear decision elements and choosing some of them to be used on recognizing each character. The main advantage of the learning procedure used in this machine is reducing the number of classes from N classes to a series of two-classes problems.

R.C.Tou (1972) designed a system for handwritten character recognition depending on feature extraction and multilevel decision making. Set of topological features are used to automatically convert the handwritten characters into computerized form and to sort them in primary classes. Local analysis is done for the characters in each primary category by a secondary stage. Testing this method on digital computer shows 6 percent misrecognition.

Wizer (1975) presented a multichannel coherent-optical correlator method for optical character recognition. Using ultrasonic light modulators a radar pulses can be compressed in order to feature the input signal. After applying the signal to an ultrasonic light modulator an optimal comparison mask can be correlated with the signal using the offset frequency method. After that a number of masks are used to recognize the feature of characters.

H.D.Grane (1977) provided an online data entry system for hand printed characters. This system uses a ball-point pen. It is dynamic because it recognizes the character depending on real-time detection and analysis instead of the detection of the complete input pattern. The recognition depends on the analysis of sequence directions that is taken by the pen while writing. This system uses the information that is taken from the pen itself not from the writing on the paper. ASCII code words are the output of the system.

Schurmann (1978) developed a multifont word recognition system for postal address reading based on multiple-channel/ multiple-choice which accepts the character images as input and classifies them into three types: capital letters, small letters or numerals. A polynomial classifier is then applied to assist in address reading. A rank-ordered list of character meanings and a channelspecific figure of confedence are the outputs of each channel. Figure of confedence is used to count the alternatives numbers in the rank-ordered list and to calculate a discriminated function for a particular word in order to distinguish between kinds of words.

comparison between the alternatives words and the legitimate words using a hash code is done. The importance of this system can be shortened in two points first moving from single character recognition to word recognition. Second using the hash code techniques. Applying this method on dictionary of 16000 entries shows a word recognition rate equal to 98 percent with 1 percent word error rate and 1 percent word rejection rate.

Peleg (1979) proposed an approach for ambiguous segmentation of handwriting. One of the main handwriting stage is the segmentation that is responsible of splitting the writing into strokes depending on some characteristics such as the height and curvature. Collected strokes then are used to set up letters then these letters are collected to create a word using a merit function. This research discussed the computation of a merit function for every combination of strokes, and select the highest maximal merit combination. The number of combinations is very large so a relaxation process is used to solve this problem by reducing the ambiguity at each node of the graph which represents the ambiguous segmentation.

2.2.2 1980's

J.Capitant (1980) discussed an application of optical character recognition techniques to bandwidth compression of facsimile data. The high quality transmission of documents is the main goal of the facsimile which depends on the elimination of redundancy of encoding technique. The combined symbol matching is considered the most efficient encoding technique. CSM detects and extracts isolated patterns from scanned documents and sends them to the receiver, and stores them in a library and compares each isolated pattern with previously encountered library patterns.

Two cases are available. First if presence of matching is detected this means that there is a recurrence of a pattern and it is already available in the receiver's library. If a misfit is found then it should be added to the library. Each library entity has a library ID which is sent instead of the pattern that leads to high compression level. Applying CSM algorithm on 86 sets of data each one containing 1,000 samples shows that no mismatches occurred and rejection just happened to badly damaged characters.

H.Yoshida (1982) developed an online handwritten character recognition algorithm for a personal computer system. This system depends on the conversion of the character pattern into single interconnected stroke pattern in order to make the recognition faster. Dynamic programming is used to recognize the characters which are written on a digitizing table.

Stefan Manke (1994) described a new standard thinning algorithm for OCR handprinted characters. Thinning is a preprocessing stage. It depends on the accuracy of data any fault in the data leads to misrecognition. Some of the data faults are loops in the skeleton and spurious tails. Standard thinning algorithm aims to reduce the rate of failure caused by these faults by using some preprocessing heuristics before applying the thinning algorithm. Testing the algorithm on a set of 689 unconstrained handprinted alphanumeric characters leads to less than 1 percent failure of the algorithm. H. E. L (1986) presented an improved version of zhang and suen algorithm which was discussed. It canceled the disadvantages through keeping the necessary structures and maintained afixed speed as the previous algorithm.

Radmilo M.Bozinovic (1989) developed system for single offline cursive script words recognition which may based on a global context or word-shape analysis. It consists of arrange of level, I-level, C-level, E-level, L-level and W-level. Input image is processed through the hierarchy of these interrelated levels. The principle of this technique is based on dividing the words into segments. Left to right scanning is done to recognize segments into letters. Bypassing the segmentation is another method in which a global word features such as ascenders, descenders and holes are needed to recognize the words.

Fang-Hsuan Cheng (1989) proposed a modified Hough transform techniques. Hough techniques are used generally for line detection because it has a powerful advantage that it can't be influenced by gaps or noise. That was the reason behind developing modified Hough transform (MHT) method based on dynamic programming matching for the Chinese characters recognition. This approach based on extracting the strokes by transforming the Chinese char acters from spatial domain to parametric domain. It is characterized by two capacities, first it considers a good technique for Chinese character recognition because the strokes of Chinese are linear and the HT treats them as lines. The second one is ease of computation which saves a lot of times comparing with the global feature approach. Applying the algorithm on 351 handwritten Chinese characters leads to rate of 49.5 percent. The technique is sensitive to rotation so this rate could be decreases if the characters are over 10 degrees rotation.

2.2.3 1990's

This period witnessed major developments in the area of banking applications and postal, portable computers. It was the beginning of the emerging of some devices such as scanners, pen, pads, electronic papers, which required the developments of asset of algorithms to deal with this changing. Some of these algorithms are dynamic programming, matching hidden markov, models (HMM) and neural networks (NN)(Jay Ashree R.Prasad, 2010).

R. O. Marc Parizeau (1990) provide a comparative analysis of three algorithms for online signature verification. the idea behind this research is to compare the performance of regional correlation, dynamic time warping and skeletal tree matching according to some criteria such as error rate, execution time and number of parameters. Regional correlation works on the principle of dividing the signal into segments and finding the best match for each paired of segments. Skeletal tree matching depends on using the distance between each signal's tree to determine the distance between signals themselves. Comparison between the three algorithms shows that, no algorithm is better than the other according to the error rate criteria. Regional correlation is faster than the other two in terms of execution time. Regional correlation has just one parameter but on the other side dynamic warping has no parameter. H. Al-Yousefi (1992) proposed statistical approach depends on the method of moments of projections for Arabic characters recognition. The algorithm consists of two steps. The first one is the segmentation of the characters into primary and secondary parts. The secondary parts are identified separately to reduce the classes from 28 to 18. The primary parts are then normalized depending on the zero order moment. Characters are then applied to 9-D feature vector. Quadratic discriminated function is used to accomplish the classification.

Krishna S. Nathan (1993) introduced a statistical approach for automatic discrete handwriting recognition using continuous parameter hidden Markov models. Handwriting is digitized on special tablet using pen-based computer devices then it is sent to a recognizer. Online recognition is done because all the handwriting temporal data is included with the sent data. To present the variation of each character a left to right hidden markov model (HMM) is used. Gaussian distributions are applied to each part of the HMM to get the output probabilities. Maximum probability and a set of feature vectors are used to execute the recognition. Series of experiments were done on both cases writer-dependent an writer-independent to compare the performance of HMM with the elastic matching algorithm. The results show that HMM is the best in both cases.

Stefan Manke (1994) applied a multi-state time delay neural networks (MSTDNN) on both online single character and cursive handwriting recognition. MS-TDNN was originally used for speech recognition. In this approach single network architecture consists of segmentation and recognition and gathers the advantages of TDNN and the dynamic time warping. Time-shift invariant is provided by time delay neural network to accomplish a high online- single character recognition rate. Recognition of continuous words can be done by combining the dynamic time warping algorithm (DT) with TDNN architecture. Words consist of a range of characters, each character is presented in a set of states which picture the first, the middle and the last part of the characters. DTW, states and input layer are the main layers of TDNN. . Each state has scores which can be computed using the TDNN.Words are recognized and modeled by a sequence of states in the DTW layer. Using DTW algorithm id an optimal alignment path for each word that can be founded. All the activations in this path are sent to the output stage. Testing the algorithm on 2000 words lead to word recognition rate up to 97%.

I.S.I.Abuhaiba (1994) presented an automatic system for cursive Arabic handwritten recognition. It depends on clustering-based skeletionization algorithm (CBSA) that can be used to get the skeleton of a pattern. The system consists of two parts the preprocessing stage and the recognition stage. Preprocessing aims to provide a general representation of the characters. Model for the character is obtained in the recognition stage using stroke schemes. Speed is the only disadvantage of this system because it needs a high speed devices and parallel processing architecture to overcome the complexity of the recognition algorithm.

Marc parizeau (1995) presented method for cursive handwriting recognition based on allograph models. Fuzzy-shape grammars are used to identify the morphological characteristics of this allograph to be used in the recognition of cursive handwriting. This approach divides the problems into two parts, recognition of graphic symbols and reading the message that is obtained from the coding of these symbols. Optimized parts are then grouped together in order to build reading system. 84% to 91% was the recognition rate that obtained from applying this method on 600 characters written by 10 writers. R.K.Powalka (1995) discussed new approach for online wholistic recognizer used in hybrid recognition system. Using diverse interactive segmentation is the main principle of this approach. Combining different recognition methods leads to significant improvement. In order to improve the wholistic recognizer which depend on ascender and desender word shape a vertical bars recognizer was developed. Vertical recognizer based on vertical bars. Features that used for word shape matching are height, number and position of these bars. Vertical bars end is presented using fuzzy sets. Representation of vertical bars is compared to all the patterns stored in database. Better recognition rate is achieved by vertical bars recognizer.

Adnan Admin (1996) has used artificial neural networks for hand-printed Arabic characters recognition. This approach contains a set pf steps, digitizing is the first one in which the original image is converted to binary one by using a 600 dpi scanner. Second step is preprocessing which aims to remove the noise that obtained from the conversion method, and to provide help in extracting some features from the binary image. Parallel thinning algorithm is used as a third step in order to thin the entire image. Fourth, detection algorithm is applied on the input image using 3x3 windows to extract some feature points. Skeleton tracing is the fifth step in which a graph representation is built by scanning the input image's skeleton from right to left. Primitives' features of Arabic characters such as dots, hamza, loops, curves and lines have to be extracted. Finally, a five layers neural network is used to classify each possible character. Recognition rate that was obtained after testing the system on 10 different users' writings is 92%. B.M.F Bushofa (1997) introduced a system for Arabic characters recognition based on structural classification. Principle of this technique is considering the whole 120 classes of Arabic characters. Proper selection in the segmentation stage is used to avoid generating of new classes. Binary image is smoothed by reducing the noise that occurred during digitizing, filling the holes, and by removing the small areas. Words are segmented into isolated characters. Secondary features of Arabic language such as hamza, madda and dots have to be removed before sending the characters to the recognition stage. Number of classes are reduced from 120 to 32 after applying the segmentation and the smoothing stages. Features are extracted and their information can be stored to be used in the final recognition stage. Recognition phase is divided into two phases, first, characters are classified using the decision tree according to the minimum number of features, any rejected character is sent to the second stage to be compared with a set of templates. Second,minimum distance criterion is used to classify the characters. Recognition rate of 92.23% is reported after testing the technique over 426 samples including four fonts.

Saleh A. Alshebeili (1997) presented technique to recognize Arabic characters using 1-D slices of the character spectrum based on estimating the projections of the characters on the X- and Y axes of the Fourier spectrum. 2-D spectrum is used to extract features, 10 of them are used as model features that generated by using distance measure criterion. Input characters are represented by the model which has minimum distances. Fourier descriptors are the criterion that used to compare this system with other recognition methods. Using 10 features of X- projection leads to 99.06% recognition rate. 99.94 % is the recognition rate if 10 Y- projection features is used. Testing a system of 10 Fourier descriptors leads to 97.5% recognition rate. The presented technique is better than this in terms of recognition rate and speed because the spectrum is calculated using fast Fourier transform while the other ones use standard equation to calculate the Fourier descriptors.

Andrew W.Senior (1998) provided a system for the recognition of offline cursive handwriting. It specially designed to deal with large vocabulary task of text transcription. It was tested over a database of writing from one writer. Preprocessing, recognition and postprocessing are the main stages of this system like any other recognition techniques.Text documents must be scanned using a scanner. Words on the scanned image are segmented. Normalization is applied to the binary image in order to remove the variations. Representation aims to form the information in the image in suitable way in order to be processed in the recognition stage. Neural network is used to estimate the recognition. HMM is used to combine the likehoods in order to find the best choice of word. After testing this algorithm an 87% recognition rate is reported.

I.S.I Abuhaiba and S.Datta (1998) presented an offline handwriting recognition based on the conversion of thinned image strokes into line approximations. Binary image is the digitized copy of original one obtained by using a scanner. Smoothing is done to the binary image. Extraction of some strokes is done, the output is smoothed binary image represents each stroke. Thinning algorithm is applied on each smoothed image. Smoothed thinned binary image is converted into straight line approximation. Enforcement of temporal information aims to extract some strokes information from straight line approximations. Tokens are produced from the segmentation of cursive strokes. In token's recognition each token is checked to find if it belongs to a class or not, if it's not then its tree data structure can be sent to learning process. Learning of token strings includes recombination of a set of tokens into meaningful set and logical strings. The main and secondary strokes are then separated. Main strokes are then ordered from right to left after extracting their lines. CSH interpretations, enumeration and requirement tree (ERT) are used to present every possible interpretation of the main strokes. Characters are then formed by combining the ERTs and the secondary strokes. Redundant secondary strokes are resulted from the character formation, they need to be manipulated in order to produce list of ordered lines of ordered lists of words. Subwords recognition rate of this system is 55.4%.

Peter Shaohua Deny (1999) introduced technique for offline handwritten signature verification. Number of features need to be identified for different signature of the same person in order to know if this signature is forgery or not. Wavelet transforms is used to decompose the curvature data into multiresolutional signals. Application of closed contour is included in this approach in order to transform an image into one dimensional signal which helps in extracting dynamic and fixed features. These features are then used to compare the signatures by calculating a threshold value to know if the comparison is accepted or rejected. Threshold can be determined for each writer by set of training samples hence it is difficult to find a fixed threshold for all writers because the variation in writing styles. Verification algorithm based on the length of closed contour is then applied to recognize the characters. A.Gillies (1999) designed a system for Arabic recognition in document images. The research focuses on low resolution and low quality images. It reads complete Arabic pages as binary Tiff images then the image enters page decomposition stage which produces a text single line. Single character is then produced from text segmentation phase. At the end segment recognition is done using a neural network classifier. 93% recognition rate is achieved when applying this system on images digitized at 200x200 dpi and 89% at 100x200 dpi.

Erik, Erl, Trenkle, and Schlosser (1999) designed a system for Arabic recognition in document images. The research focuses on low resolution and low quality images. It reads complete Arabic pages as binary Tiff images then the image enters page decomposition stage which produces a text single line. Single character is then produced from text segmentation phase. At the end segment recognition is done using a neural network classifier. 93% recognition rate is achieved when applying this system on images digitized at 200x200 dpi and 89% at 100x200.

2.2.4 2000 onward

This era saw a major development in the recognition process because of the development that took place in the field of postprocessing which helped to improve the overall efficiency of the system. Combination of several independent recognizers and the use of some dictionaries or lexicons of the language models are some of the most famous techniques that used in this era(Jay Ashree R.Prasad, 2010).

M.A.Ismail (2000) have proposed a new approach for off-line Arabic signature recognition and verification. This approach consists of two main phases, recognition and verification which are based on a defined grade of membership in a set of signatures. Removing the noise from the signature image while keeping the characteristics of the signatures helps in feature extraction and classification. Recognition phase is divided into two steps. First, feature extraction which aims to extract the common features that differentiate one class from the others. Second, Classification in which a multistage classifier is used where a pre classification is done on a group of similar signatures. In order to solve individual identification within a group a recognition scheme is needed. Last stage of classification based on the threshold of the candidate class to determine if the sample is not recognized. Verification is the second phase of this approach that is based on fuzzy concepts used to verify the signer identity by using the information that obtained while signing. The results of testing of this approach show a recognition rate of 95% and verification rate of 98%.

Beatrice Lazzerini (2000) in their work have proposed a linguistic fuzzy recognizer to be used in offline handwritten characters recognition based on the classification of the characters shape where each character is presented by reference models. The character image is divided into fuzzy partitions on both horizontal and vertical axes in order to generate a linguistic representation of the character. Each character has a linguistic reference that is obtained from the linguistic representations of the character training set. Defined weighted distance is used to compare the linguistic representation of any new character with the linguistic reference models of each character. Recognition of the character depends on the amount of congruence between the reference models and the character representation. Testing this approach on the NIST handwritten database shows a recognition rate of 69.5%.

M.Dehghan (2001) have developed a holistic approach for handwritten Arabic word recognition. This kind of approaches depend on the size of the available lexicon.If the size of the vocabulary is small It is much better to be used in some application such as the recognition of the legal amount in bank cheques. This research focuses on the recognition of handwritten city names using a model discriminant discrete HMM. Scanner with 300 dpi is used to scan an image of postal envelope. A range from 1 to 198 of appropriate labels is assigned to the city names that are extracted from the image. Preprocessing consists of fourth steps first, binarazation which aims to binarized the image using a predefined threshold. Second, noise removal that depends on using a morphological closing operation followed by a morphological opening operation to eliminate the spurious segments of the binarized images. Slope correction and baseline estimation and stroke width estimation are the third and fourth steps of preprocessing. Images must be converted to suitable form to be used by the HMM recognizer module. Depending on the image contour some features are extracted using a sliding window of right to left scanning. Self-Organizing vector quantization is used for smoothing of the probability distributions of HMM. The recognition rate of this system is 65%. Kye Kyung Kim (2001) developed a system for the recognition of handwritten numeral strings using combined segmentation method. Recognition systems are composed of three modules presegmentation, segmentation and recognition. This system based on decision value generator which has a confidence value that is used in the presegmentation module as a basis to classify the numeral strings into sub images (Isolated digits). Four kinds of segmentation points are used in segmentation module to segment the isolated digits depending on a segmentation- free method. Finally, the type of the digit is detected using confidence value. Touching digits are segmented using reliability value. 96.7% recognition rate is obtained after applying the system on unknown length numerals of NISTSD.

Toufik Sari (2002) presented a new algorithm (ACSA) for Arabic character segmentation. Segmentation is one of the most important steps in the character recognition. Any error occurs in this process causes an error in the process of Identifying the character. Arabic character segmentation algorithm (ACSA) is divided into five steps: first, preprocessing which aims to binarize and smooth the image using a threshold value in order to extract some features. Second, connected components extraction where 8-connected contour following algorithm is used to smooth the contours while traveling from one pixel to another in the eight different directions. Sequence of X-Y coordinates of the outer contour of the word is the output of this algorithm. Third, feature extraction, MF-list is the output of this phase which is used to solve some extracted morphological features. Fourth, topological filtering operation (TOFO), In this phase a segmentation points is determined by primitive analysis of the outer contour of Arabic word. Fifth, contour dissection, this phase aims to solve the slant correction problem by extracting the lower and the upper outer contour of the word image and adding its secondary parts and loops to reconstruct the body of the character. In this proposed system a RECAM recognition system is used for recognizing the segmented Arabic characters depending on four independent tree layer neural networks. Satisfactory results were obtained from testing this algorithm on small database of some isolated Arabic characters.

Velissarios G. Gezerlis (2002) developed a new system for offline optical recognition of the Hellenic Orthodox Byzantine music (OBMR). After scanning the document using 300 dpi scanner, the scanned images enters in three stages. First, segmentation stage which aims to get the bitmaps of the extracted characters. Second, recognition stage which consists of three steps: preprocessing, feature extraction and classification. Identification numbers of the characters are the output of this stage. Third, the semantic musical group recognition stage, the input of this stage is the character Identification numbers (Cids) the output is the group identification numbers (Gids). Finally the fourth stage is the postprocessing which aims to produce the conversion of a true type font, or the performance of the BM. Three structural features were developed: the Euler number, the principal an X direction and the ratio of the horizontal bounding rectangle of the character. These three structural features are used to divide the set of 71 character classes into 19 smaller subsets. In order to get an efficient classification system. A discrete wavelet transform (DWT) is applied on the computed 4- projections of the characters in order to form feature vectors. The over all accuracy of this system is 99.4%.

Sofien Touj (2003) used a Hough transform technique for Arabic optical charC acter recognition. Hough transform (HT) is good for detecting alignment, ascenders and descanters in an image. Defining a mapping between an image space and a parameter space is the basic principle of the (HT). The proposed approach uses the GHT which is an extension of (HT). The Idea of this technique is to store all the edge pixels of the target image in a table and defining a reference points for each position. R- Table is then built for each model of the character. Recognition of the characters based on segmentation / recognition approach. Testing this approach over a 166873 samples of Arabic characters shows a recognition rate of about 97%.

Mohammad Sarfaz (2003) presented a new technique for the recognition of Arabic printed text using artificial neural networks. The proposed system involves five stages: first, Image a acquisition aims to get a digitized image of the text using scanning system. Second, preprocessing in which the image is enhanced by doing some drift correction and removal of isolated pixels. Third, stage is the segmentation consists of 4- level process that leads to segment the text into individual characters. Feature extraction is the fourth stage which aims to extract the features of the character and compare it with the existing ones in order to identify the characters. Using 7 moments invariant descriptors neural classifiers are in the last stage (recognition) to classify the characters according to their computed modified moment invariants. Testing this system shows a recognition rate of 73%.

Somaya Alma'Adeed (2004) developed a new schema for Arabic handwritten recognition based on multiple hidden Markov models (HHM). As any recognition this system consists of particular steps. Preprocessing, recognition and postprocessing. Normalization is done in the in the preprocessing stage in order to get a reliable presentation of the handwritten word. Recognition is carried out by following a series of steps. The original lexicon is reduced using global feature engine. Local HMM classifier is applied to these reduced lexicons. A vector quantization (VQ) method is used to estimate the data likelihoods for each frame. There are two basic properties that distinguish this system from others. First one is its ability to deal with similar Arabic word with different meanings. Second one is using some of the properties of the position of features in the character to provide a recognition help to VQ AND HMM classifiers. Applying this system over a database of 100 writers shows a recognition rate of 60% without using a post processing.

M. Z. Saeed Mozaffari Karim Faez (2005) proposed a technique for Farsi / Arabic handwritten Zip Code recognition based on statistical method embedded with statistical features. Skeleton of the numerals which is obtained after the smoothing operations is partitioned into primitives using some standard feature points. These primitives include special information that describes the topological structure of the skeleton. Direction and curvature of the skeleton is statistically described by calculating the average and variance of X and Y changes in each primitive. PCA algorithm is used to normalize the different lengths of the global codes. Recognition of each numeral can be achieved by using a nearest neighbor classifier (NNC). 94.44% is the classification rate that is obtained after applying this approach on test sets of various people with different educational background and different ages.

Ramy El-Hajj (2005) employed an analytical approach for offline handwriting recognition using 1D HMM. The main principle behind this approach is defining a subset of baseline dependent features using lower and upper parts. Sliding window is used to extract two types of features, distribution features and concavity feature. Characters are then classified using HMM -based classifier that doesn't need a pre-segmentation process. IFN/ENIT database was used to evaluate this system. It consists of 26459 handwritten words of 946 Tunisian town/villages names. Evaluation test of this approach shows a recognition rate of 74.9%.

Nadir Farah (2006) presented a system for Arabic checks literal amount recognition based on classifiers combination and syntax analysis. This approach depends on global approach which uses structural high-level features such as ascenders, descenders and loops. Three modules classifiers are used multilayer neural, k-nearest neighbor and fuzzy k-nearest neighbor classifier. These three classifiers are used in parallel to classify the extracted features. Neural network uses a supervised training. K-NN aims to search among the training set using thresholding to reject or accept a class. Fuzzy K-NN is used to compute the membership of the tested word which is then assigned to the highest membership class. Different schemas are then used to combine the results of the classifiers. Syntactic analyzer takes the candidate words as in put to find the best matches. The proposed approach shows a recognition rate of 96% which depends on the performance of the classifiers.

Roman Bertolami (2006) has investigated rejection strategies for offline handwritten text line recognition. Deciding whether to accept or reject recognized word is depending on different confidence meatures. Three strategies of confidence meatures are used. The first one depends on knowing how many times a recognized word appears in the candidates set. Second strategy is based on the class of the word because some words are recognized more than others. Third one is used to combine the output of the candidate sources by using a multilayer perception. A HMM model is used in the recognition phase experiments on the IAM database shows an acceptance rate of 20%and a rejection rate of less than 19%.

Jari Hannuksela (2006) proposed an online handwritten recognition for mobile interaction. This approach depends on collecting and estimating the interface motions of the device using discrete cosine transform where the sequence of each handwritten character is stored in a form of (x,y) coordinates. Motion trajectories are then resembled in order to calculate the DCT based features. K-nearest neighbors (K-NN) is used to classify the characters. 92% to 98% is the range of the recognition rate that was achieved in experiments. F.Menasri (2007) in their research has described shaped-based alphabet for off-line Arabic handwriting recognition. Letter-body alphabet is the new shape alphabet that was developed in this research. It explores some prior knowledge of Arabic properties and redundancies in order to reduce the number of classes. The main thing that distinguishes this system from others is the combination between the dots/ diacritics recognizer with the letter-body recognizer which makes it more suitable for large vocabulary application. Combination of neural network and HMM is used for recognition purposes. Recognition rate of 87% was achieved from testing this approach on the IFN/ENIT benchmark database.

V. M. H. E.-A. Saeed Mozaffari Karim Faez (2008) discussed the lexicon reduction for offline farsi/Arabic handwritten word recognition using dots concepts. The main principle of this technique is to eliminate unlikely candidates by extracting and representing the number and the position of dots with respect to the baseline from the input image. Recognition rate and recognition time are effected by the lexicon size.recognition speed is the most important criterion if the lexicons are large. On the other hand, recognition accuracy is critical issue for small lexicons. There are a variety of methods can be followed to minimize the size of lexicon such as knowing some information about the application environment characteristics of input pattern, and the clustering of the same lexicon entries. After extracting the dots candidates a classifier is needed to categorize them into classes (single dot, double dots, and triple dots). A model discriminate discrete HMM is used for recognition because this approach is used for reading 200 city names from postal address fields. Testing this system leads to a recognition rate of 73.61% with a 93% lexicon reduction.

Safwan Washah (2009) proposed a new system for segmentation of offline handwritten Arabic words based on contour and skeleton segmentation. Different styles of handwriting and connectivity of Arabic letters make the segmentation of Arabic handwritten a challengeable task. The proposed system aims to present a new segmentation algorithm which segments the word into maximum five segments per character called maximum segments per character (MSC). Maximum character per segment (MCS) is the definition of any segment contains up to three letters. This way helps to minimize the lexicon size and make the problem simple. A segmentation rate of 93% was achieved after testing the algorithm over 6300 words from 45 different documents written by 18 writers.

Hanan Aljuaid (2009) in their research have proposed a complete system of off-line Arabic handwriting recognition based on projection profile and genetic approach. Genetic algorithm is computer science technique specially used for optimization and search problems. Preprocessing in this approach is done using a thinning algorithm to thin the image word, and then to extract the vertical and horizontal projection profiles. In order to define the shape of the characters number of features such as, length and width of the character, loops, hamza and points are needed to be extracted by tracing the boundary of the image from right to left. Feature extraction is followed by recognition phase based on genetic algorithm which depends mainly on fitness function that is calculated for each input vector in order to choose the best fitness one.

37350 characters were used to test the system. Experiments show a recognition rate of 87% with rejection rate of 8.4%. Touru Wakahara (2010) addressed the problem of improving the recognition accuracy by developing a new technique to extract, describe and evaluate handwritten deformation in a deterministic, parametric manner. Two basic ways were described. First, using a 2D warping to generate a handwriting deformation vector field (DVF). This method is concerned to solve optimization problems by finding the optimal matching between an input and target images in grayscale. Second, merging the DVF by a parametric deformation model of global/ local offline transformation in order to extract only the loworder deformation component. IPTP CDROM1B numeral database was used to evaluate the proposed technique. The results show that discriminating un-natural deformation from natural one using decomposition of the DVF improves the recognition accuracy to a recognition rate of 92.1% which is better than 87.0% that obtained using 2D warping.

C.Kermorvant (2010) described the improvement of recognition performance by parallel combination of different recognizers based on different features and models. The basic structure of this system is composed of the combination of three recognizers, GausianHMMbased on sliding window and a hybrid neural network based on grapheme segmentation. Rank based combination methods are used to combine these recognizers depending on scores. Each one of these classifier shows a recognition rate of 80% individually. A 90% recogC nition rate is achieved after the combination process. Combining classifiers that based on the same technologies is less efficient than the combination of different classifiers. 83.6% is the recognition rate that was obtained from the combination of two classifiers based on sliding window HMM. On the other hand combining two classifier based on a sliding window and grapheme respectively yield recognition rate of 89.1%.

BRIJESH VERMA (2011) presented a new segmentation technique for cursive handwritten words based on binary segmentation and multiple confidence values. This technique consider as a repetition of binary segmentation and fusion of segment confidence. Segment confidence-based binary segmentation (SCBS) is mainly depends on a set of suspicious segmentation points (SSPs) which are generated using an over segmentation technique. In order to know if any improvement was made the binary segmentation was applied on each SSP an evaluation of segment confidence is done. A segmentation path which is used to divide a segment into two parts is found using a binary segmentation. A presegment is the name of a set of segments before applying the binary segmentation. A postsegment is the name of a set of segments after applying the segmentation. A highest fusion point is calculated for each segmentation point and compared to the presegments if it is higher than the presegments the segmentation point consider as the final point. If it's not no more segmentation is needed. Benchmark database was used to evaluate this technique.

Salvador Espana-Boquera (2011) described the recognition of unconstrained off-line handwritten texts using hybrid hidden Markov (HMM) and artificial neural network (ANN) models. The main principle of this technique is using artificial neural network (ANNs) in preprocessing stage to remove the slant and slope from text lines and to normalize the size of the images. The slope and the horizontal alignment are estimated using local extrema from a text image. Normalization is achieved by computing the reference lines of the slope and slant-connected text. Recognition is based on hybrid HMM/ANN where graphemes is modeled using left to right markov chains and single neural network is used to estimate the emission probabilities.