Ensemble learning using multi-objective optimisation for arabic handwritten words

• Main Author: Ghadhban, Haitham Qutaiba
• Format: Thesis
• Language: English; English; English
• Published: 2021
• Subjects: T Technology (General)
• Online Access: http://eprints.uthm.edu.my/8421/1/24p%20HAITHAM%20QUTAIBA%20GHADHBAN.pdf; http://eprints.uthm.edu.my/8421/2/HAITHAM%20QUTAIBA%20GHADHBAN%20COPYRIGHT%20DECLARATION.pdf; http://eprints.uthm.edu.my/8421/3/HAITHAM%20QUTAIBA%20GHADHBAN%20WATERMARK.pdf; http://eprints.uthm.edu.my/8421/

Arabic handwriting recognition is a dynamic and stimulating field of study within pattern recognition. This system plays quite a significant part in today's global environment. It is a widespread and computationally costly function due to cursive writing, a massive number of words, and writing style. Based on the literature, the existing features lack data supportive techniques and building geometric features. Most ensemble learning approaches are based on the assumption of linear combination, which is not valid due to differences in data types. Also, the existing approaches of classifier generation do not support decision-making for selecting the most suitable classifier, and it requires enabling multi-objective optimisation to handle these differences in data types. In this thesis, new type of feature for handwriting using Segments Interpolation (SI) to find the best fitting line in each of the windows with a model for finding the best operating point window size for SI features. Multi-Objective Ensemble Oriented (MOEO) formulated to control the classifier topology and provide feedback support for changing the classifiers' topology and weights based on the extension of Non-dominated Sorting Genetic Algorithm (NSGA-II). It is designated as the Random Subset based Parents Selection (RSPS-NSGA-II) to handle neurons and accuracy. Evaluation metrics from two perspectives classification and Multiobjective optimization. The experimental design based on two subsets of the IFN/ENIT database. The first one consists of 10 classes (C10) and 22 classes (C22). The features were tested with Support Vector Machine (SVM) and Extreme Learning Machine (ELM). This work improved due to the SI feature. SI shows a significant result with SVM with 88.53% for C22. RSPS for C10 at k=2 achieved 91% accuracy with fewer neurons than NSGA-II, and for C22 at k=10, accuracy has been increased 81% compared to NSGA-II 78%. Future work may consider introducing more features to the system, applying them to other languages, and integrating it with sequence learning for more accuracy.