In recent years, there was intensive research on Arabic Optical Character Recognition (OCR), especially the recognition of scanned, offline, machine-printed documents. However, Arabic OCR results are unsatisfactory and are still an evolving research area. Exploring the best feature extraction techniques and selecting an appropriate classification algorithm lead to superior recognition accuracy and low computational overhead. This paper presents a new Arabic OCR approach by integrating both of Extreme Learning Machine (ELM) and Non-dominated Rank Sorting Genetic Algorithm (NRSGA) in a unified framework with the aim of enhancing recognition accuracy. ELM is adopted as a neural network classifier that has a short processing time and avoids many difficulties faced by gradient-based learning methods such as learning epochs and local minima. NSRGA is utilized as a feature selection algorithm that has better convergence and spread of solutions. NSRGA emphasizes ranking among the solutions of the same front, along with elite preservation mechanism, and ensuring diversity through the nearest neighbor method reduces the run-time complexity using the simple principle of space-time trade-off. The Experimental results reveal the efficiency of the proposed model and demonstrated that the features selection approach increases the accuracy of the recognition process.