A multi-filter based license plate localization and recognition framework

Abstract

A license plate is regarded as the unique identification of a vehicle, which makes the license plate recognition (LPR) an indispensable operation in intelligent transportation systems (ITS). Since many techniques related to the LPR are restricted to specific working conditions, a multi-filter based LPR framework for the plate localization and the character recognition is proposed to solve the issues. In the localization phase, chromatic and morphologic filters are cooperated with each other under flexible criterions to detect candidate plate regions accurately. Plate characteristics, such as the length-to-width ratio, the size of a character, etc. are utilized in the character segmentation phase. In the recognition phase, a back propagation (BP) neural network is trained for the character recognition. 800 images taken from various scenes under different conditions are used to evaluate the accuracy of the proposed framework. The experimental results show that the missing rate of localization is close to zero, and the accuracy of the plate localization and the recognition is 98.4% and 93.8% respectively. Moreover, the overall accuracy of the multi-filter based framework is 93.1%.