Bangla License Plate Recognition from Motion Blurred Images using Deep Learning

Research Supervisor
Journal Title
Journal Title
Journal ISSN
Volume Title
We propose an idea which is de-blurring the motion blurred images of Bangla license plates. Our AI-based system is capable of recognizing the speedy vehicles that have motion-blurred license plate numbers. Our system will be able to detect license plates from speedy vehicles. Mostly, the system will benefit to prevent traffic law violations, such as ‘hit and run’ and speed limit breaks. To accomplish it, we have built a smart AI-based system which includes license plate segmentation model, motion deblur model and OCR model. The dataset that we are using to train the models contains about 150K photos of Bangladeshi vehicles and number plates where majority of the photos are gathered by ourselves. We have applied YOLO v5 model to build our segmentation model to detect motion blurred license plates from speedy vehicle images. The segmentation model has near about 90% accuracy. For motion de-blur, we are implementing the idea of CNN and Auto-encoder. Both of the models were fairly stable to deblur the license plates, with the training accuracy of 71% and 66%. Though the second model has 66% training accuracy, but it works faster and gives better output than the other model. For the first model, 8 out 10 random samples give decent result for not more than 35% motion blur and for the second model, 8 out of 10 random samples give adequate result up to 45% motion blur from various angles. We are approaching the second model as our final model. Moreover, we are currently assessing existing neural networks along with our own models and unique dataset for novelty.
Department Name
Electrical and Computer Engineering
North South University
Printed Thesis