The Designing of a Small Statistical Feature Set for Effective Handwritten Character/Digit Recognition

In this paper, a new feature extraction method based on statistical features of input image is devised which applicable to both binary and grayscale images and named asSpread Direction. This feature set is a 27 dimensional vector which extracts three statistical features from nine region ofinput image. The number of this region is selected such that the correlation between features is minimized. Experiments have been applied to MNIST and HODA datasets that are handwritten digit datasets in English and Persian languages, respectively. The proposed method has been compared to other well-known feature extraction techniques with different classifiers. Experimental results show that classificationaccuracy of proposed method is higher than almost all othermethods. On MNIST dataset, Spread Direction obtains 96.42% recognition rate and On HODA dataset, obtains 97.63% recognition rate.

کلیدواژه ها:

feature extraction ، handwritten digit ، spread direction ، preprocessing ، classification