Number Plate Recognition

Multiple Artificial Neural Network Number Plate Recognition

Aurthor: Matthew Vlietstra

ICT219 Intelligent Systems Student, Murdoch University

Email: Use the Contact form (Removed to prevent spam).

Date: 24/10/2011

Abstract

The purpose of this report is to outline the details of the construction, setup, and results of a Multiple Artificial Neural Network (MANN1), and a Single Artificial Neural Network (SANN2) Number Plate Recognition program. The purpose of using two different setups is to determine which ANN (Artificial Neural Network) delivers better results. The program utilizes various image processing techniques3, which results in the extraction of characters in a numberplate. These characters are then procedurally inputted into a neural network, producing a result of each individual character. These results are then combined, forming a digital representation of a number plate4.

Statistical information in the form of graphs are used to illustrated the detection rate of MANNs against SANNs. Results indicate that the MANN was able to detect 8.07% (Average) more characters and 10.76% (Average) number plates then the SANN.

The neural networks are batch tested using 48 categories of pseudo number plates (20,592 number plates, totalling 2.16GBs); each category has its own unique effect. A majority of these effects do not appear/ are different from the learning data. Results of these tests are illustrated using column graphs.

Keywords: Multiple Artificial Neural Network, Number Plate Recognition, OCR, Optical Character Recognition, comparison of a multiple neural network and a single neural network

1. Uses two neural networks, one for numbers and one for letters. A pattern is used to determine which neural network to use, i.e. “naaannn” – the n represents a number, the ‘a’ represents alphabet characters. This method could, in theory produce better results (as it the neural network won’t get confused between 1’s and I’s or 0’s and O’s). Appendix 13 illustrates an abstract MANN

2. Conventional ANN, uses a single neural network for letters and numbers – Appendix 14 illustrates an abstract SANN (ANN – Artificial Neural Network).

3. These include

   a. Binary Images – an image that can only have two possible values for each pixel ( 0 or 1)

   b. Image Cropping – this involves isolating a specific part of an image and discarding the rest.

4. Appendix 1 provides an abstract illustration of the number plate detection process.

The Main Form