Home > Research > Publications & Outputs > Vehicle plate recognition using improved neocog...
View graph of relations

Vehicle plate recognition using improved neocognitron neural network

Research output: Contribution in Book/Report/Proceedings - With ISBN/ISSNConference contribution/Paperpeer-review

Published
Publication date2013
Host publicationArtificial Neural Networks and Machine Learning – ICANN 2013: 23rd International Conference on Artificial Neural Networks Sofia, Bulgaria, September 10-13, 2013. Proceedings
EditorsValeri Mladenov, Petia Koprinkova-Hristova, Günther Palm, Alessandro E. P. Villa, Bruno Appollini, Nikola Kasabov
Place of PublicationBerlin
PublisherSpringer Verlag
Pages628-640
Number of pages13
ISBN (Electronic)9783642407284
ISBN (Print)9783642407277
<mark>Original language</mark>English

Publication series

NameLecture Notes in Computer Science
PublisherSpringer
Volume8131
ISSN (Print)0302-9743

Abstract

This paper describes a novel vehicle plate recognition algorithm based on text detection and improved neocognitron neural network, similar to [1] and based on Fukushima’s neocognitron. The proposed recognition algorithm allows us to improve the recognition speed and accuracy comparing to both traditional neocognitron and some state-of-art algorithms (multilayer perceptron, topological methods). It can be used as a solution for image classification and analysis tasks. As an example, the neocognitron can be utilized for symbols recognition [2]. We propose several modifications comparing to the Fukushima’s modification of the neocognitron: namely, layer dimensions adjustment, threshold function and connection Gaussian kernel parameters estimation. The patterns’ width and height are taken into account independently in order to improve the recognition of patterns of slightly different dimensions. The learning and recognition calculations are performed as FFT convolutions in order to overcome the complexity of the neocognitron output calculations. The algorithm was tested on low-resolution (360 ×288) video sequences and gave more accurate results comparing to the state-of-the-art methods for low-resolution test set.