Automatic Vehicle Detection and Classification System using Advanced Deep Learning Architectures

Abstract

This dissertation aims to develop an automatic vehicle detection and classification system that leverages advanced deep learning architectures. The system comprises three parts. The first part employs advanced convolutional neural networks (CNN) to classify images of five vehicle types in Thailand. By comparing nine different CNN models with data augmentation, we found that MobileNets is the best method in terms of accuracy, speed, and size. The second part uses ensemble methods to combine multiple CNN models to recognize vehicle type and make (logo) using a "partial training set" technique. This approach improves accuracy and reduces overall runtime. In the third part, we propose a hybrid structure of a generative adversarial network (GAN) and a CNN for object detection using YOLO technique to recognize Thai license plates. By testing different GAN architectures and YOLO networks, we found that the hybrid of ESRGAN-YOLOv7 outperformed other combinations in terms of accuracy. Overall, this dissertation provides a comprehensive solution to the problem of automatic vehicle detection and classification using the latest deep learning methods, highlighting the importance of using ensemble methods and partial training sets to improve accuracy and reduce runtime. The proposed system has the potential to be utilized in various real-world applications, including video surveillance systems and mobile devices.

-