Optimizing Handwritten Numeral Recognition for English and Devanagari Using MNIST and CPAR Data

In this paper, a technique for the categorization of offline handwritten digits in the Devanagari script as well as the Roman script (English numbers) is presented. Finding out how to get the best possible recognition results while concurrently working with several scripts is the primary purpose of this study. The approach that has been described makes use of an easy profiling method for the extraction of features, together with an innovative implementation of Linear Discriminant Analysis (LDA), and a neural network architecture that has been adapted specifically for numeral classification. An exhaustive series of trials including 36,000 examples of handwritten numerals were carried out as part of the evaluation of the effectiveness of this approach. These samples were chosen at random from the CPAR datasets, with 22,000 being put to use in the construction of the training set and the remaining 14,000 being assigned to the role of the test set. In addition, the well-known MNIST database was used, which provided 60,000 samples for training purposes and 10,000 samples specifically for testing. It was determined via thorough testing that the LDA classifier produced variable results. These varying results are essential in emphasising the success of the method as well as possible areas for improvement for the method. According to the results, this strategy, which combines profile-based feature extraction with advanced classification algorithms, has the potential to lead to major breakthroughs in the area of handwritten number identification.