A Neural Network-based Software Tool

for Number-Plate Recognition

A. Frosini, M. Gori*, L. Pistolesi

Dipartimento di Sistemi e Informatica - Universit? di Firenze
Via di Santa Marta 3, 50139 Firenze, Italy
* E-mail : [email protected]

Abstract

In this paper we describe a software tool running on standard platforms, including PC and several Unix Workstations for the number-plate recognition. The software is based on a hybrid model where a society of autoassociator-based neural networks are properly coordinated by modules charged of segmenting the number-plate and the single characters, respectively.

1. Introduction

In the last few years, the availability of impressive computational resources at affordable cost has stimulated the interest for several unconventional problems of information processing. A remarkable number of these problems are in the field of image processing, where one is mostly concerned with the extraction of information from pixel-based arrays. The problem of the recognition of the number-plate addressed in this paper has been receiving a growing attention in a number of different applications. When dealing with natural images, the problem is made particularly difficult because of the different environmental conditions that give rise to significantly different images taken from the same number-plate. A remarkable noise is introduced by the the light conditions and by the variable angle between the camera and the moving car. Basically the rectangle can have different dimensions, the colors, the fonts, and the characters' thickness can be different. All these problems make number-plate recognition difficult and, sometimes very hard, also for humans. In addition, the automatic approach of the problem cannot neglect computational contraints that are necessary in order to develop any commercial product.
In this paper, we give an overview of the solutions that we have adopted for approaching the problem and discuss the experimental results found in motorway toll environment. The recognition is based on a hybrid model that acts using a sort of hypothesize and verify technique. The extraction of the number-plate and the character segmentation are based on classical edge detection techniques while the character recognition is carried out by Multilayer Perceptrons (MLP) acting as autoassociators.