ETHIOPIAN CAR LICENSE PLATE RECOGNITION USING DEEP LEARNING

The focus of this research is to develop a system that assist humans in reading car
license plate. Such a study is important as the number of traffic on roads becomes
increasing constantly, the manual process in car license plate recognition becomes a
serious problem for traffic management system which not only detect and track a
vehicle but also identify it. Initially a dataset that contains 930 car images was prepared
for model comparison purpose. Two object detection algorithms (Faster R-CNN and
SSD) were trained and tested on the same dataset using the same model to select the
best candidate. The metrics for the comparison were accuracy, average prediction time,
and total training time taken. It was found that Faster R-CNN gives high accuracy, short
average prediction time, and short total training time. After that additional car and
cropped license plate images were added to the prepared dataset and based on this, two
object detection networks were trained using Faster R-CNN one for plate detection and
another for character recognition on the detected plate. The proposed approach has been
tested on test set and later collected images of national license plate of Ethiopia. Both
the trained models were achieved a high accuracy which is 99 and 98.89 mAP over 0.5
IoU for plate detection and character recognition respectively and takes on average 12s
to complete the recognition of a license plate. The study could be further investigated
on other countries.