ABSTRACT
This research developed a license plate and classification scheme using deep learning architecture which utilized transfer learning using pre-trained Convolutional Neural Network (CNN). The developed scheme used images obtained from Caltech dataset, Peking University VehicleID (PKU VehicleID) dataset and a developed dataset of vehicle licence plates from Ahmadu Bello University, Zaria called the ABU dataset. De-noising, downscaling operation and grayscale conversion was applied on the acquired image to reduce the cost inqured in using the original image. Sobel operation was performed to detect the edge of the pre-processed image. Edge density filtering and connected component analysis were used to extract and verify the region which constituted the licence plate number. AlexNet model pre-trained on ImageNet was used to extract features and classify the detected license plate candidate's regions. The performance of the developed scheme was evaluated on 150 images of each dataset of PKU VehicleID, Caltech, and ABU test images taken under different conditions. With the Caltech dataset the scheme achieved a precision rate of 85.10%, recall rate of 98.50%, and recognition accuracy of 98.08% while the PKU VehicleID dataset gave precision rate of 97.91%, recall rate of 97.91%, and accuracy of 100%. For the ABU dataset, the method obtained a precision rate of 95.8%, recall rate of 100%, recognition accuracy of 99.82%. The results for the developed deep learning-based scheme showed some performance improvements of 3.96% and 14.75% in the precision and recall rate, respectively, and 8.15% improvement in recognition rate, when compared with the existing scheme which utilized the edged based approach with SVM and achieved detection rate of 98% on the PKU VehicleID dataset, 90% on the Caltech dataset, and 96% on the ABU dataset in the presence of complex backgrounds and highly variable license plate patterns.