A Pen Based Intelligent System for Educating Arabic Handwriting

The main objective of this thesis is to build an intelligent tutor system to
recognize and diagnose children handwritten mistakes. Handwritten recognition system has been an important area due to its applications in several
fields such as optical character recognition, license Plate recognition, etc.
This work is focusing on the recognition as a part of handwritten Arabic
characters and digits recognition that faces several challenges, including the
unlimited variation in human handwriting and the large public databases. The
proposed system was designed and developed for automatically recognizing
handwritten Arabic characters and digits using deep learning architectures.
The system also detects and diagnoses Arabic preschool children handwritten
characters based on immediate feedbacks. We show how to automate this task,
creating a system that checks the handwriting errors such as stroke sequence
errors, stroke direction error, stroke position error, and extra stroke error.
For Arabic handwritten digits recognition , this work presents a deep
learning architecture called Convolutional Neural Network (CNN). LeNet-5
model was implemented by two convolutional, two pooling layers ,and two
classification layers. CNN based on LeNet-5 was used to train and test the
MADBase database (Arabic handwritten digits images) that contains 10k
testing images and 60k training images. The CNN based LeNet-5 outcomes
an average accuracy 88%. We improve the performance of the classification
using unsupervised learning approach called Stacked AutoEncoder (SAE)
for Arabic handwritten digits categorization. SAE shows that the use of
SAE leads to significant improvements across different machine-learning
viii
classification algorithms. Stacked autoencoder is giving an average accuracy of 98.5%. We improve the performance of handwritten digits using a
convolutional neural network (CNN). The optimization methods were implemented to increase the performance of CNN. A comparison is held amongst
the results, and it is shown by the end that the use of convolutional neural
network led to significant improvements across different machine learning
classification algorithms. Convolutional neural nework was trained and tested
on MADBase database. The proposed CNN approach is describe as follows
: INPUT → CONV → RELU → POOL → CONV → RELU → POOL →
FC → RELU → FC → Output. Moreover, the improved convolutional neural
network is giving an average recognition accuracy of 99.15%.
For Arabic handwritten characters recognition , this work presents an
unsupervised learning approach using SAE designed for Arabic handwritten
characters categorization. SAE was trained and tested using our proposed
dataset that contains 3360 testing images and 13440 training images. We
show that the use of SAE leads to bad classification outcomes. SAE is giving
an average accuracy of 64%. This thesis presents a supervised learning
approach using convolutional neural network. The convolutional neural
nework was trained and tested using our database that contains 16800 dataset.
The proposed CNN approach is described as follows : INPUT → CONV →
RELU → POOL → CONV → RELU → POOL → FC → RELU → FC →
Output. Moreover, the convolutional neural network is giving an average
recognition accuracy of 94.85%.
Finally, an intelligent tutor application was built for Arabic preschool
children called Intelligent Arab Teaching. Intelligent Arab Teaching system
allows Arab children to practice at anytime and anywhere. The Intelligent
Arab Teaching provides a useful feedback to Arab children to correct their
mistakes. Experimental results indicate that the proposed intelligent system
successfully detect handwritten strokes errors with immediate feedback.