Arabic script is one of the most widely used scripts in the world. Besides 340 millions native speakers, more than 200 million nonnative speakers are there in the world. Many languages such as Urdu, Persian, etc. adopted this script. Therefore, recognition of optical handwritten Arabic characters has received significant attention from the past decade. Unlike numerals of any other scripts, some Arabic numerals have more than one representation. For instance, Arabic ‘4’, ‘5’ and ‘7’ have two variations each. Considering these variants as the same class may deviate the recognition performance. The present work proposed a mechanism to deal with such variations for improved classification. At the first phase, the recognition problem is considered as a 13 class classification problem instead of 10. Then, in the second phase, the classes are reorganized and post-processed for improved classification. A comparative study has also been included with the conventional approach that considers the variants as the same class. Experimental results reflect the efficiency of the proposed technique.