The use of invariant moments in hand-written character recognition

TL;DR

This paper introduces a neural network-based approach for handwritten character recognition that leverages invariant moments and a specialized preprocessing step, showing competitive results against traditional methods.

Contribution

It presents a novel neural network architecture with built-in knowledge and a preprocessing technique based on invariant moments for improved recognition accuracy.

Findings

Neural network achieves high recognition accuracy.

Preprocessing effectively groups characters by components.

Comparison shows competitive performance with KNN.

Abstract

The goal of this paper is to present the implementation of a Radial Basis Function neural network with built-in knowledge to recognize hand-written characters. The neural network includes in its architecture gates controlled by an attraction/repulsion system of coefficients. These coefficients are derived from a preprocessing stage which groups the characters according to their ascendant, central, or descendent components. The neural network is trained using data from invariant moment functions. Results are compared with those obtained using a K nearest neighbor method on the same moment data.