DEVELOPMENT OF AN ALGORITHM FOR PERSIAN HANDWRITING DIGITS RECOGNITION BASED ON MLP AND PNN CLASSIFIERS AND USING CLUSTER CENTERS

Pattern recognition is a branch of machine learning that recognizes the patterns and regularities in a set of data, and digit recognition is considered one of the pattern recognition categories. Due to the similarities between some digits in each language, especially in Persian, different algorithms have been developed to recognize the handwriting digits with the least error and in the shortest time complexity. One of the most commonly used methods in data classication is the neural network algorithm. While neural networks have been used in the literature for handwriting digits recognition, the combination of clustering approaches and neural network classiers has not been considered for this problem. Accordingly, this paper proposes an algorithm based on the combination of clustering approaches and neural network classiers to recognize the Persian handwritten digits accurately. This algorithm performs pattern training and recognition based on Probabilistic Neural Networks (PNN) and multilayer perceptron (MLP) neural networks. In this regard, after extracting the characteristic loci feature and zoning from each image in the training database, the data of each of the ten classes has been clustered using linkage, Partition Around Medoids (PAM), and Fuzzy C-Means (FCM) methods based on the extracted features. Then, the new ten classes resulting from the clustering algorithm are taught by one of the two classiers, including MLP and PNN. In order to determine the optimal number of clusters in each class, the Tabu search optimization algorithm, one of the most accurate meta-heuristic optimization algorithms, is used. The performance of the proposed algorithms is evaluated and compared with existing algorithms based on the HODA dataset. Based on the results, the proposed algorithm accurately recognizes the Persian handwritten digits. In addition, the proposed method performs more accurately and much faster than most competing algorithms.