| Summary: | License plate detection is the process of locating a license plate in each image. There exist multiple
datasets of license plates from different countries. However, license plates in these datasets are different
from those on Singapore’s vehicles. License plate in Singapore comes in the form of different shapes,
size, colour, and font. There also exist different kinds of object detection models from one-stage
detectors to two-stage detectors.
Thus, this project proposes the use of YOLOv4, a one-stage detector model to identify license plates on
Singapore’s vehicles. The use of different sets of data has been experimented with to find the best performing model for this use case.
Three different models are proposed based on different datasets. The first model ‘model 1’ is trained on
images gathered from Open Image Dataset v6, which consists of foreign license plates to gain baseline
results on the model when predicting local license plates. Next, ‘model 2’ includes a mix of images
from Open Image Dataset v6 as well as self-taken and annotated images to understand the difference in
results by including images from Singapore’s vehicles. Lastly, ‘model 3’ aims to further improve the
dataset by performing data augmentation which helps the model to detect license plates in images where
they could be obscured or orientated making the input image considered difficult.
The model ‘model 3’ has achieved the best results in terms of accuracy and detection speed compared
to the other models.
|
|---|