Stacking ensemble model of deep learning and its application to Persian/Arabic handwritten digits recognition

One of the challenges in recognizing handwritten texts is the individual style of writing. There is the structural similarity of the different digits to each other in writing. Along with the mentioned challenge, these similarities in the text may increase and make it difficult to correctly recognize digits and numbers. In this paper, a new model for recognizing handwritten digits is presented. The proposed model is a stacking ensemble classifier. This classifier is based on the convolutional neural network (CNN) and the bidirectional long-short term memory (BLSTM). Another innovation of the model is the use of the probability vector of images class as the input of the meta-classifier layer. One of the strengths of BLSTM is the ability to learn arrays and vectors; therefore, from a technical point of view, considering the output probability vector of the first model as the input of the meta-classifier (BLSTM) improves the accuracy of the deep learning model. The reason for using stacking ensemble classification is the sameness of the main body of some Persian/Arabic digits (e.g., "2, 3, and 4"). Also, the style the author writes makes classes that are not similar in a structure similar to each other, which causes errors incorrect recognition. This model helps to recognize the correct set of input digits by examining the structure of similarities. To achieve a reliable result in the face of this challenge, this model has been tested on a large Persian/Arabic dataset to cover a wide range of writing styles from different individuals. The dataset has a total of 102,352 data which 60,000 of them are for training data and 20,000 of them are for test data in ten classes of digits that are used in this paper. The result of using this database is to improve the recognition performance of these challenging digits. In examining the dataset presented by the model, the accuracy rate of the training set was 99.98%. And the sample accuracy rate in the test set was 99.39%. That is, compared to experimenting with the convolutional neural network and other researches, the rate has increased. (All codes available at http://web.nit.ac.ir/-h.omranpour/.) (C) 2021 Elsevier B.V. All rights reserved.