A Novel Domain-Specific Feature Extraction Scheme for Arabic Handwritten Digits Recognition

The most crucial step for the success of any character recognition system is the extraction of good features. A wide variety of well-known universal feature sets were used in character recognition problems; such as gradient, Kirsch, and contour features. This paper argues that a domain-specific features extracted based on their ability to characterize the classes at hand in a way similar to how humans intuitively discriminate among those classes can outperform universal features set. Arabic handwritten digits recognition problem is used to prove the superiority of domain specific features over universal ones. Results show that a carefully chosen feature vector of only 35 features could outperform many universal feature sets of hundreds of features in both recognition accuracy and speed.