Feature extraction with an improved scale-invariants for deformation digits

Abstract

This paper presents an alternative formulation on improved scaled-invariants using higher order centralized moments for digits with deformations. We claim that deformation digits would be digits with improper shapes, unconstrained styles of writing and different orientations. A detail experimental evaluation of the utilizing various moments order as pattern features in recognition of handprinted and handwritten digits have been carried out using the proposed invariants. We use scale-invariants of centralized moments of order 2 for the numerator and order 4 for the denominator while preserving the scale factor of the same order. Unconstrained digits are rotated clockwise and counter clockwise of 45 degree. As a comparison, we generate geometric moment invariants on these digits. We train these invariants using standard back-propagation and modified backpropagation in the classifications phase. We found that the results are promising with an improved scaled-invariants of higher order, and the classifications of the digits are successfully recognized.

Keywords:

• Handwritten digits
• Feature extraction
• Geometric moment invariants
• Backpropagation

C.R. Categories:

• I 4.7
• I 5.1