Cursive Arabic Handwriting Recognition System Without Explicit Segmentation Based on Hidden Markov Models

In this paper we present a system for offline recognition cursive Arabic handwritten text which is analytical without explicit segmentation based onHidden Markov Models (HMMs). Extraction features preceded by baseline estimation are statistical and geometric to integrate both the peculiarities of the text and the pixel distribution characteristics in the word image. These features are modelled using hidden Markov models. The HMM-based classifiercontains a training module and a recognition module. The training module estimates theparameters of each of the character HMMs uses the Baum-Welchalgorithm. In the recognition phase, feature vectors extracted from an image are passed to a network of word lexicon entries formed of character models. The character sequence providing the maximumlikelihood identifies the recognized entry. If required, the recognition can generate N best output hypotheses rather than just the single best one. To determine the best outputhypotheses, the Viterbi algorithm is used.The experiments on images of the benchmark IFN/ENIT database show that the proposed system improves recognition.


I INTRODUCTION
The recognition of cursive Arabic handwriting is an active area of pattern recognition research. The variability of words, letter shapes are context sensitive, inter and intra word spaces, thecursive nature of Arabic handwriting, the skew and slant ofcharacters and words makes the construction of offlinerecognition system a challenging task. Researches have tried various approaches for text recognition employing various techniques for pre-processing, featuresextraction and classification [A. Lawgali 2015].
The subject of this article concerns the recognition of cursiveArabic handwriting [M. T Parvez 2013] [AL-Shatnawi 2011].Several systems are available based on two approaches; a globalapproach that considers the word as non-divisible base entityavoiding the segmentation process and its problems.This approach is reliable and applicable for vocabularies oflimited size. Against, the analytical approach is based on thedecomposition of the word sequence into characters orgraphemes proceeding by a segmentation phase. The latter canbe explicitly based on a priori division of the image into subunits(letters or grapheme) or implicitly based on a recognitionengine to validate and rank the segmentation hypothesis.The approach used in our system is analytical based on implicitsegmentation; segmentation and recognition are carried outjointly.
The first step of a handwriting recognition system afterpreprocessing is the extraction features. The objective of thisphase is the selection of primitives relevant for the next steps ofclassification and recognition. The performance of a recognitionhandwritten system largely depends on the quality and therelevance of the extracted features. In our system after thebaselines estimation, the extracted features are statistics actingon the densities of pixels and structural extracted from therepresentation of the character shapes. Hidden Markov models (HMMs) are used for classification [S. Azeem 2013] [AlKhateeb 2011][ A. Maqqor 2014]. There aremany reasons for success of HMMs in text recognition includingavoidance of the need to explicitly segmentation. In addition,HMMs have sound mathematical and theorical foundations.Each word is described by a model built by concatenating themodels of the component character. The system performstraining and recognition of words and characters. The remainder of this paper is organized as follow. Section 2presents a detailed description of the features extraction precededby baselines estimation. Section 3 is focused on classificationstep. The performance of therecognition system has been experimented on the benchmarkdatabase IFN/ENIT and the obtained experimental results areshown and analysed in section 4. The paper finally concludeswith some conclusions and perspectives.

Baseline Estimation
The goal is to find, for a given word, the positions of the two following parallel lines (  The approach used is based on the horizontal projection curvethat is computed with respect to the horizontal pixel density,knowing that the skew and slant correction of words are made inpreprocessing step to harmonize the course of the slidingwindows in the extraction features. LB corresponds to the maximum of the projection profile curve,then, the algorithm scans the image from top to bottom to findthe upper baseline, which corresponds to the first line with aprojection value higher or equal to the average row density.Thus, the handwritten variability of word in both zones, upperand lower are considered.

1.2Extraction Features
The features extraction method used In each window we extract a set of 28 features represent thedistribution features based on foreground pixels densities andconcavity features. Each window is divided into a fixed numbern of cells. Some of these features are extracted from specificareas of the image delimited by the word baselines. In our experimentation the parameters are set to n = 20 cells andthe width = 8 pixels. This leads to a total of Nf= 28 to calculatein each frame.
Let:  n(i) : the number of foreground pixels in cell i  r(j) : the number of foreground pixels in the jth row of a frame.
The extracted features are the following: f1: density of foreground (black) pixels. f21,…, f28 : represent the density of foreground pixels in each vertical column in a frame.
In each frame 28 features vector are extracted, these features are statistical and geometric to integrate both the peculiarities of the text and the pixel distribution characteristics in the word image, which capture the type of strokes (curved, oriented, vertical, and horizontal).

2.1Hidden Markov Models
A Hidden Markov Model (HMM) is a doubly stochastic processwith an under-lined stochastic process (Markov chain) that is notobservable (it is hidden), but can only be observed throughanother set of stochastic processes that produce the sequence ofobserved symbols .In order to define an HMM completely, following elements areneeded, λ = (N, M

2.2Character and word models
The used approach is analytical and based on character modelingby HMM. Each character model has a right-left topology andparameters: number of hidden states, state transition probabilitiesand observation probabilities. There is no specific theory to set these parameters, so thesolution is empirical. Many considerations can be taken intoaccount in setting these parameters and in particular thetechnique used in the generation of sequences of observations. Inour system we used a model with four states for each characterwith three transitions for each state (Figure 3).

IV. EXPERIMENTATIONS AND RESULTS
In order to investigate the potential of our systemfor offline cursive handwriting recognition, the benchmarkdatabase IFN/ENIT is used [Peschwitz 2003], that contains atotal of 26459 handwritten words of 946 Tunisian town/villagesnames written by different writers.We used the toolbox HTK (Hidden Markov Model Toolkit[S.Young 2006]) to model the characters and words.
The table below shows the experimental results of our systemcompared to other recognition systems using the samebenchmarking database IFN/ENIT, divided into four sets a, b, cfor training and d for testing :

V. CONCLUSION & PERSPECTIVES
In this paper, we present a recognition system of Arabic cursivehandwriting based on hidden Markovmodels. The extracted features are based on the densities offoreground pixels, concavity and derivative features usingsliding window, some of these features depends on baselinesestimation. The modelling proposed has improved recognition,and shown encouraging results to be perfect later.Many points are yet to be achieved, firstly modeling a characterallows deformations related to its context (next and previouscharacter). To account possible deformations, contextualmodeling of characters is opted. The word is no longer seen as asuccession of independent characters, but as a sequence ofcharacters in context. Word models are the concatenation ofcontext-dependent characters models: the trigraphe, thismodelling will allow building more accurate and more efficientmodels. Taking into account the characters environment allowsmore precise and more effective models to be built. However,this implies a multiplication of HMM parameters to be learned, itwould be the focus of our next work. Then language models willbe incorporated to refine and improve the results and lead to amore efficient system.