Encoder-Embedded Feature Enhancement in Convolutional Neural Networks for Arabic Handwritten Recognition
Oussama Alkayed, Marwa Amara, Nadia Smairi, Abdelmalek Zidouri · Procedia Computer Science · 2024
In the field of Arabic handwriting recognition, the search for models that perfectly combine efficiency and accuracy is still ongoing. This paper introduces a novel approach that harnesses the synergy between autoencoders and convolutional neural networks (CNNs) to set a new benchmark in the recognition of Arabic handwritten characters. Our method centers around an autoencoder that meticulously learns a compact representation of the characters, followed by the integration of its encoder into a CNN architecture, dubbed the Encoder-CNN. The prowess of our model is demonstrated through rigorous experiments on the Arabic Handwritten Characters Dataset (AHCD), where it achieved a best accuracy of 98.87%. These results not only underscore the model’s ability to capture the intricate nuances of Arabic script but also its robustness in generalizing to unseen data.