Ensemble of adaboost cascades of 3L-LBPs classifiers for license plates detection with low quality images

Abstract

Due to the plate formats and multiform outdoor illumination conditions during the image acquisition phase, it is challenging to find effective license plate detection (LPD) method. This paper aims to develop a new detection method for identifying vehicle license plates under low quality images using image processing techniques. In this research, a robust method using a large number of AdaBoost cascades with three levels pre-processing local binary patterns classifiers (3L-LBPs) are used to detect license plates (LPs) regions. The method achieves a very high accuracy for detecting LP number from one vehicle image. The proposed method was tested and trained with the images from 630 and 400 vehicles, respectively. The images involve many difficult conditions, such as low/high contrast, dusk, dirt, fogy, and distortion problems. The experimental results demonstrate very satisfactory performance for LP detection in term of speed and accuracy, and were better than the most of the existing methods. The processing time for the whole testing LPD system was about 1.63 seconds to 2 seconds. The overall probability detection, precision, and f-measurement are 98.56%, 95.9% and 97.19%, respectively; with false positive rate 5.6%.