Libyan vehicle license plate detection and recognition using radila basis function

Abstract

An integrated vehicle plate detection and recognition system generally aims to detect the license plate (LP) and recognize its characters. The process basically includes LP detection, LP extraction, character segmentation, feature extraction and recognition. Due to its wide range of applications such as traffic management, security etc., and this topic is intensively researched especially in the field of image processing. Furthermore, the differences in systems, colours, backgrounds, foregrounds, font and style of the license plates from one country to another add more problems and challenges for new researches. Although various detection and recognition methods that have been proposed in the previous studies, their performances especially in terms of accuracy are fallen short, and room for improvements is still wide open. Thus, this thesis presents an integrated approach for detecting and recognizing Libyan license plates based on Radial Basis Function Neural Network (RBFNN). The method begins with the preprocessing of the image using edge detection and morphological operations. In the detection stage, connected component analysis is used to locate unique objects, from which the unwanted objects are removed using the filtering process. Geometric and Global features are used to prepare the identified objects before their classification as Plate and non- Plate using RBFNN. In the recognition process, for character segmentation, a simple template is derived to extract and differentiate digits and Arabic words, as the Arabic word is not segmented into individual letters like digits. The outputs are improved using median filtering and connected component analysis. Statistical and structural features are used in feature extraction, while the classification is performed using RBFNN. The performance of the proposed method is evaluated and tested using 200 frontal images of Libyan national license plates. Experimental results have shown that the proposed method has produced convincing results with accuracy rates of 93% and 91% for detection and recognition, respectively, and has also outperformed the colour-based approach.