Licence plate detection and recognition

The main goal of the project is to make a model that will reveal the license plate on the picture of the car and read the digits from it. We will do this using two libraries; Object Detection to find a Region of Interest and Easy OCR to extract the text

Dataset

The data set contains 433 images (10% used for test set) and bounding box annotations of the car license plates (xml files).

Step 1

The first step is to detect and locate the license plate on the input image.

Step 2

Then we read the label from the detected license plate using the OCR tool.

Detection

Transfer learning

Tensorflow Object Detection API using pre-trained model

ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8

Trained on COCO dataset


Results

Results are mostly accurate with some mistakes on small plates and multi lined texts.

This is one of the accurate readings.



In this example number "7" is read as the letter "z".


Problems with multi-lined licence plates:

Future work

Detection:
    - Parameter tuning
    - Training

OCR:
    - Homography
    - Morphological operations

Presentation

Project presentation can be downloaded here.