Automatic Check Digits Recognition for Arabic Using Multi-Scale Features, HMM and SVM Classifiers

We propose in this work two Automatic Arabic (Indian) digits recognition systems using a real-life dataset of 3000 bank checks. The systems extracts features from training-set images of 7390 isolated digits (0-9). These features are multi-scale in which they capture narrow, intermediate, and large-scale qualities of the image. The gradient features correspond to the narrow scale, the structural features correspond to the intermediate scale, and the concavity features correspond to the large-scale. These features are employed by two different statistical classifiers; Hidden Markov Models (HMM) and Support Vector Machines (SVM). The two independent recognition systems utilize the proficient CENPARMI Arabic bank check database for training and testing. In order to select the optimal parameters for feature extraction and for the HMM classifier, the CENPARMI training dataset is divided into training and verification subsets. After adapting the two systems’ parameters, they are tested on unobserved 3035 digit images. The average recognition rates for the HMM and SVM systems are 97.86% and 99.04%, respectively. The presented systems provides state-of-the-art recognition results on the CENPARMI database, as they reported a higher recognition rates when compared to twelve previously published systems, especially for the SVM system. After analyzing the classification errors, the authors conclude that some of these errors are inevitable as they are most probably attributed to errors in labeling the original database, distinct writing styles of certain digits, and genuine faults.