Efficiency of Dropout Regularization in Character Recognition: Introducing the Dropout Efficiency Score Within Intelligent Systems Architectures

Arabic script, characterized by its cursive and context-sensitive nature, presents unique challenges in handwriting recognition. While Convolutional Neural Networks (CNNs) offer potential solutions, they often deal with overfitting, especially on challenging or limited datasets. To counter this, dropout-a prominent regularization technique-has been employed. This research delves into the impact of dropout on CNNs customized for AHR. We embark on an extensive experimental exploration, modulating dropout rates, layer placements, and CNN configurations, leveraging benchmark datasets for Arabic handwriting as our evaluation playground. Recognizing the need for a powerful measure we propose the Dropout Efficiency Score (DES), a key metric assessing the balance between model performance and computational overhead, central to our research contributions. The knowledge acquired highlights the crucial role of dropout in enhancing CNNs’ generalization for AHR. Nonetheless, the efficacy varies, contingent on architectural choices, model depth, and the intricacies of the Arabic script. This investigation results in an innovative concept that guides the effective deployment of dropout in CNNs, achieving a balance between maximum output and improved efficiency.