Abdel Ennaji

Semi-continuous HMMs with explicit state duration for unconstrained Arabic word modeling and recognition

In this paper, we describe an off-line unconstrained handwritten Arabic word recognition system based on segmentation-free approach and semi-continuous hidden Markov models (SCHMMs) with explicit state duration. Character durations play a significant part in the recognition of cursive handwriting. The duration information is still mostly disregarded in HMM-based automatic cursive handwriting recognizers due to the fact that HMMs are deficient in modeling character durations properly. We will show experimentally that explicit state duration modeling in the SCHMM framework can significantly improve the discriminating capacity of the SCHMMs to deal with very difficult pattern recognition tasks such as unconstrained handwritten Arabic recognition. In order to carry out the letter and word model training and recognition more efficiently, we propose a new version of the Viterbi algorithm taking into account explicit state duration modeling. Three distributions (Gamma, Gauss and Poisson) for the explicit state duration modeling have been used and a comparison between them has been reported. To perform word recognition, the described system uses an original sliding window approach based on vertical projection histogram analysis of the word and extracts a new pertinent set of statistical and structural features from the word image. Several experiments have been performed using the IFN/ENIT benchmark database and the best recognition performances achieved by our system outperform those reported recently on the same database.

Using diversity in classifier set selection for arabic handwritten recognition

The first observation concerning Arabian manuscript reveals the complexity of the task, especially for the used classifiers ensemble. One of the most important steps in the design of a multi-classifier system (MCS), is the its components choice (classifiers). This step is very important to the overall MCS performance since the combination of a set of identical classifiers will not outperform the individual members. To select the best classifier set from a pool of classifiers, the classifier diversity is the most important property to be considered. The aim of this paper is to study Arabic handwriting recognition using MCS optimization based on diversity measures. The first approach selects the best classifier subset from large classifiers set taking into account different diversity measures. The second one chooses among the classifier set the one with the best performance and adds it to the selected classifiers subset. The performance in our approach is calculated using three diversity measures based on correlation between errors. On two database sets using 9 different classifiers, we then test the effect of using the criterion to be optimized (diversity measures,), and fusion methods (voting, weighted voting and Behavior Knowledge Space). The experimental results presented are encouraging and open other perspectives in the classifiers selection field especially speaking for Arabic Handwritten word recognition.

Optimizing a Neural Network Architecture with an Adaptive Parameter Genetic Algorithm

This article deals with the use of genetic algorithms to optimize the architecture of a neural network. After a brief recall of our original neural network (named Yprel network), we show that a simulated-annealing-like technique has been advantageously replaced by genetic operators. Indeed, tests on character recognition (NIST handwritten database) have shown that the generalization rate has been improved, the mean network size has been reduced by a factor 3 and the learning speed has been significantly increased. Moreover, a portable adaptive mutation probability has been introduced which enables a parameter-free learning.

New dynamic classifiers selection approach for handwritten recognition

In this paper a new approach based on dynamic selection of ensembles of classifiers is discussed to improve handwritten recognition system. For pattern classification, dynamic ensemble learning methods explore the use of different classifiers for different samples, therefore, may get better generalization ability than static ensemble learning methods. Our proposed DECS-LR algorithm (Dynamic Ensemble of Classifiers Selection by Local Reliability) enriched the selection criterion by incorporating a new Local-Reliability measure and chooses the most confident ensemble of classifiers to label each test sample dynamically. Confidence level is estimated by proposed reliability measure using confusion matrix constructed during training level. After validation with voting and weighted voting fusion methods, ten different classifiers and three benchmarks, we show experimentally that choosing classifiers ensemble dynamically taking into account the proposed L-Reliability measure leads to increase recognition rate for Handwritten recognition system using three benchmarks.

MLP Modular Versus YPREL Classifiers

We present two connectionist modular approaches which are potentially able to deal with real applications as their size does not increase drastically with the size of the problem. The first model relics on a very simple cooperation of modular MLP networks specially designed for some sub-tasks. The second is based on a new methodology using a particular processing element (“neuron”) called yprel. The main characteristics of the approach are: (i) An yprel classifier is a set of yprel nets, each net being associated to a particular class; (ii) the learning is supervised and conducted class by class; (iii) the structure of the net is not a priori chosen, but is determined step by step during the learning process. Both approaches are compared on a well-known classification task (recognition of typographic characters) in terms of performance rates.

Signature Verification for Offline Skilled Forgeries Using Textural Features

This study explores the effectiveness of two textural measurements on signature verification for skilled forgeries. These texture features include 2D autoregressive coefficients and run-length distributions. Signature images corresponding to 521 writers from the GPDS960 database were used to evaluate the performance of these features. Comparison of the proposed textural features with a number of state-of-the-art features realized interesting results. The run-length features out perform other features for a sufficient number of genuine signatures in the training dataset.

Classification and Feature Selection by a Self-Organizing Neural Network

This article describes recent improvements of an original neural network building method which could be applied in the particular case of 2 input neurones. After a brief recall of the main building principles of a neural net, authors introduce the capability for a neurone to receive more than 2 inputs. Two problems then arise: how to chose the input number of a neurone, and what becomes of the decision rule of a neurone? Treating these problems leads to an original feature selection method (based on genetic algorithms) and leads to adapt a linear discrimination algorithm to non separable problems. Experimental results for a handwritten digit recognition problem confirm the efficiency of the method.

A Distributed Classifier Based on Yprel Networks Cooperation

In this paper we present a scheme of classification based on a particular processing element (“neuron”) called Yprel. The main characteristics of the approach are: (i) an Yprel classifier is a set of Yprels networks, each network being associated with a particular class; (ii) the learning is supervised and conducted class by class; (iii) the structure of the network is not a priori chosen, but is determined step by step during the learning process; (iv) the learning process is incremental: each network improves its own learning base with the errors of the previous test; (v) networks cooperate: each network can use the outputs of the previously builded networks. Preliminary results are given on a well-known classification task (recognition of typographic characters).

Clustering data: dealing with high density variations

This paper focuses on the problem of cluster analysis when data present high variations of density. The proposed method is based upon a hierarchical clustering and enables one to determine the clusters without any assumption on their number nor their statistical distribution. This method is used to design an efficient distributed neural classifier which reveal a good generalization behavior on a real problem of handwriting digit recognition (NIST database).

Spot Words in Printed Historical Arabic Documents

Libraries contain huge amounts of arabic printed historical documents which cannot be available on-line because they do not have a searchable index. The word spotting idea has previously been suggested as a solution to create indexes for such a collecton of documents by matching word images. In this paper we present a word spotting method for arabic printed historical document. We start with word segmentation using run length smoothing algorithm. The description of the features selected to represent the words images is given afterwards. Elastic Dynamic Time Warping is used for matching the features of the two words. This method was tested on the arabic historical printed document database of Moroccan National Library.