MACHINE LEARNING-BASED APPROACH FOR NIGERIAN LICENSE PLATE CHARACTER RECOGNITION

Abstract

Vehicle License Plate location and character recognition has numerous applications among which are vehicle monitoring, control of access to restricted areas, and parking lot management. This research work established three models capable of recognizing Nigerian vehicle license plate characters. The objectives are to design three models for Nigerian vehicle license plate character recognition using image processing techniques of grayscale conversion of image, adaptive threshold, connected components analysis, vertical and horizontal vehicle plate segmentation character segmentation, and character recognition. The segmented vehicle license plate characters were normalized to 2010 before feature extraction, which was carried out using Gray Level Cooccurrence Matrix (GLCM). The extracted features were used to form a feature vector for each vehicle license plate character. A class label was assigned to each character feature which forms the input to the Artificial Neural Network (ANN) and Support Vector Machine classifiers used in this research work. The first model, Template Matching, is based on correlation of vehicle license plate character with template images stored in an array. In order to improve the classification rate of this technique, pattern analysis was implemented. The other two machine Learning models: Support Vector Machines (SVM) and Artificial Neural Network classifiers were also implemented for character recognition. The dataset (1811 characters from 230 vehicle plate images) were divided into training and testing sets. The result shows that out of 230 vehicle plate images, 228 were correctly detected (99.1% accuracy), 227 vehicle plates were correctly segmented out of 228 (99.6% accuracy). Template matching model (without pattern analysis) gave a recognition rate of 90.8%. With pattern analysis, this increased to 93%. SVM (one-versus-one) model gave a recognition rate of 98.9% with 70% of data for training and 30% data for testing. However, with 65% training data and 35% testing data, the recognition rate is 98.1%. ANN model gave a recognition rate of 93.4% (70% data for training and 30% for testing). With 65% data for training and 35% data for testing, the recognition rate is 92.3% with 500 epochs and the learning rate set at 0.3. SVM model showed the best classification rate, followed by ANN model and then the Template matching model. The results shows that the proposed system performs well on the test datasets. Thus, the system is good for recognition of vehicle license plate on Nigerian roads.