RECOGNITION OF HANDWRITTEN ARABIC DIGITS IN WORST-CASE SCENARIOS

Abstract

In general, character recognition focuses on average recognition performances. In this study, we aim to maximize the probability of correct classification of handwritten Arabic digits in worst-case scenarios. Here, a worst-case scenario refers to a digit that is poorly written compared to the typical form of its category. Besides focusing on the worst-case rate, this paper also highlights the recognition of Arabic digits which are less explored in the literature in contrast to the Latin or Chinese digits. For these experiments: first, we will build minimax dictionaries of Arabic digits from the training dataset obtained from the MADBase (Modified Arabic Digits dataBase). For comparison purposes, we also train SVD (singular value decomposition) dictionaries from the same database. Each digit represents a class, thus we have 10 classes of training examples and test examples. Then, using these learned dictionaries (minimax and SVD), we evaluate the recognition rate in worst-case scenarios. The experiments are executed 100 times to allow random permutation between the samples. Results show that in most cases, the minimax approach performs better in recognizing the poorly handwritten Arabic digits.