Robust Efficient Accurate License Plate and Character Detection System Based on Simplified CNN

Zekai Fei, Dr. Meng Han, and Dr. Selena He, Department of Computer Science, Kennesaw State University, 1000 Chastain Road, Kennesaw, GA, 30144

Current license plate recognition systems struggle with image noise reduction and license plate feature detecting processes. The most common approach to achieve license plate detection is to reduce image noise and emphasize the plate features by image processing. However, such an approach is not flexible for a huge amount of datasets. With 500 image inputs, the accuracy of successfully detecting license plates with image processing is around 63%. We present an efficient and highly accurate license plate detection and character detection program based on the YOLO neural network, a simplified CNN-based neural network frame for robust image processing systems. Unlike most approaches, the system we proposed provides a method to evaluate potential noises inside dataset images so that program implementations could be more effective and more targeted to design and optimize with YOLO neural network. We evaluate our program with specified data groups, such as high/low-resolution images, tilted images, images with different noise backgrounds, etc. With such an evaluation, we proved that our program could dominate with tilted images, low-quality images, and so on. With our presented system, license plate detection accuracy improves from .63, which is performed by traditional image processing methods, to .903. Other than accuracy, our program also performs with better efficiency than traditional image processing methods. Our program generally halves the average processing time for each image.

Additional Abstract Information

Presenter: Zekai Fei

Institution: Kennesaw State University

Type: Oral

Subject: Computer Science

Status: Approved

Time and Location

Session: Oral 3
Date/Time: Mon 4:30pm-5:30pm
Session Number: 345
List other presenters in this same room and session