An automated license plate detection and recognition system based on wavelet decomposition and CNN

Abstract

The License Plate detection and recognition (LPDR) is a challenging task that plays a significant role in intelligent transportation systems (ITS). Where it could be used as a core in various applications, such as security, traffic control, and electronic payment systems (e.g. freeway toll payment and parking fee payment). A variety of algorithms are developed for this work and each one has advantages and disadvantages for extracting plates in images under different circumstances. However, the complexity of some methods requires a high calculation cost and this could be time-consuming. In the current paper, a simple and efficient method is proposed to tackle the issue of license plate detection and character recognition. The license plate is detected first based on the two-dimensional wavelet transform to extract the vertical edges of the input image. The high density of vertical edges is computed first to detect the potential areas of the license plate. Then these potential areas are verified by using a plate/non-plate CNN classifier. After the license plate is detected, the characters are segmented by using a simple method that is based on the empty distance between the characters. Finally, these character candidates are classified by training another CNN classifier. The experiments were done on vehicles that carry Moroccan license plates and showed high accuracy, where the results obtained go up to 99.43% in term of localization and 98.9% in term of recognition. Besides, the efficiency and the high accuracy of the proposed method were proved by performing a comparison with other works from the literature on different datasets. All processes of the proposed method were implemented on a Hardware Processor System (HPS) located in a VEEK-MT2S provided by TERASIC.