Improved low-cost recognition system for handwritten Bengali numerals

Handwriting recognition is very important due to its numerous potential applications. This paper is concerned about the low-cost features extraction for the development of an improved Bengali handwritten numeral recognition system. Each numeral was first resampled to a binary image of fixed size. A set of new features based on shape analysis was derived from the resampled image, and a multilayer neural network was trained using the extracted features. The recognition accuracy of the developed system was tested on both training and test sets of a publicly available Bengali handwritten numerals database at three different resolutions. Besides accuracy, the reliability of the system was also estimated using Cohen's kappa. The highest accuracy, 99.12% with reliability about 99%, was obtained for the test database at resolution of 32×32. The use of PCA reduces feature dimension from 142 to 68 resulting in a slight reduction in accuracy to 98.80%.