Recurrence sorting method for improved accuracy of unconstrained fast-moving vehicle license plate recognition system

Abstract

Automatic Real-Time Vehicle License Plate recognition systemsface a wide variety of challenges to accurately reading the number sequences present in license plates when deployed in the real world. The uncertainty of the real-world license plate recognition system is addressed in this research and proposes a novel method of filtering out inaccurate results based on recurrence-based sorting of the best readable character sequences. The problems of unconstrained real-world real-time license plate recognition have been discussed thoroughly throughout the paper and redefined the problem statement towards attaining the best readability instead of best visibility. An end-to-end cascade neural network architecture has been developed to capture license plate readings in real time. Then the novel recurrence-based iteration method was introduced to sort the Top 1 of the best reading of a unique license plate tracked across the frame where character sequence reading accuracy has been improved by up to 54% when combined with the introduced garbage factor filtering method. The experimental evidence indicated that the traditional confidence-based sorting is prone to failure due to unconstrained real-world uncertainties and accuracy is thoroughly compared with our novel method to illustrate novelty. The system has been deployed in the real world on an embedded system with a substantial amount of vehicle traffic for testing and validation.