International Journal of Electrical and Computer Engineering (IJECE) Vol. 13, No. 3, June 2023, pp. 3010~3018 ISSN: 2088-8708, DOI: 10.11591/ijece.v13i3.pp3010-3018  3010 Journal homepage: http://ijece.iaescore.com An optimized deep learning model for optical character recognition applications Sinan Q. Salih 1 , Ahmed L. Khalaf 1 , Nuha Sami Mohsin 2 , Saadya Fahad Jabbar 2 1 Department of Communication Engineering Technology Department, College of Information Technology, Imam Ja’afar Al-Sadiq University, Baghdad, Iraq 2 College of Education, Ibn Rushed for Human Science, University of Baghdad, Baghdad, Iraq Article Info ABSTRACT Article history: Received Jul 14, 2022 Revised Oct 5, 2022 Accepted Dec 2, 2022 The convolutional neural networks (CNN) are among the most utilized neural networks in various applications, including deep learning. In recent years, the continuing extension of CNN into increasingly complicated domains has made its training process more difficult. Thus, researchers adopted optimized hybrid algorithms to address this problem. In this work, a novel chaotic black hole algorithm-based approach was created for the training of CNN to optimize its performance via avoidance of entrapment in the local minima. The logistic chaotic map was used to initialize the population instead of using the uniform distribution. The proposed training algorithm was developed based on a specific benchmark problem for optical character recognition applications; the proposed method was evaluated for performance in terms of computational accuracy, convergence analysis, and cost. Keywords: Chaotic maps Convolutional neural network Training algorithm Computational intelligence Black hole algorithm This is an open access article under the CC BY-SA license. Corresponding Author: Sinan Q. Salih Department of Communication Engineering Technology, College of Information Technology, Imam Ja’afar Al -Sadiq University Baghdad, Iraq Email: sinan.salih@sadiq.edu.iq 1. INTRODUCTION Every simple artificial neural network (ANN) is basically made up of the input and output layers of neurons, but the necessity for an intermediate hidden layer in ANN gave rise to the concept of deep learning (DL) [1], [2]. So, DL can be considered a more complicated version of ANN that relies on the use of numerous layers with nonlinear processing units; most DL frameworks rely on the supervised or unsupervised data representations learning concept [2]. The year 1965 witnessed the introduction of the first deep feedforward multilayer perceptrons-based working algorithm [3] and since then, DL has improved and been adopted in many applications. DL is appliable in several areas, such as pattern recognition, neural networks, optimization, graphical modeling, artificial intelligence, and signal processing [4]–[6]. Recently, convolutional neural networks (CNN) was developed as a way of achieving a better accuracy during recognition tasks. However, one of CNN main problems is the difficult design of its architecture for specific task. Thus, several architectures of CNN have been developed, such as the LeNet architecture which was originally developed by LeCun [7], [8]. This architecture was implemented for optical character recognition (OCR) and character recognition in several documents. ConvNet is another CNN architecture that relies on the use of 7 layers where each layer has a specific role. Using the same architecture for several tasks has a poor chance of reaching optimal performance; consequently, distinct CNN architectures are built for specific tasks, which requires a lot of research work since there are many types of