license_plate_recognition
Project for Image Processing Course. The goal of the project is to recognize polish license plate from an image.
Requirements about images:
- the angle between horizontal surface and license plate is +- 45 degrees
- longer edge of license plate is greater than 1/3 of image's width
- license plate has 7 characters
- images can have diffrent resolutions
Requriments about project:
- written in Python 3.7 using OpenCV
- there's an option to use other libraries (like scikit-image) but you cannot use external OCR modules or trained models that can read characters
- maximum time for each image processing is 2 seconds
Results on private test set:
- Bounding box accuracy of license plate: 96,42% (27/28)
- Bounding box accuracy of characters in license plate: 91,32% (179/196)
- Total score: 86,42% (Every correct read character on license plate -> 1 point. Correct read of whole license plate equals to 1 point for each character + 3 additional points.)
Results on final, unknown test set:
- Total score: 63.70% (Every correct read character on license plate -> 1 point. Correct read of whole license plate equals to 1 point for each character + 3 additional points.)
- execution time per image: 0.11s
Usage
- Create virtual environment using
python3 -m venv venv
- Install requirements from
requirements.txt
- Run script with
python main.py <path_to_folder_images> <path_to_json_results_file>
How it works?
Preprocess the image.
- convert to gray scale
- resize the image for faster processing
- blur image and find image edges.
Find potential vertices of license plate using edge image.
- find contours on the image
- create bounding boxes from contours
- skip bounding boxes that doesn't match license plate height to width ratio
- get corners of potential license plate
Get bird's eye view for every potential license plate
- construct destination points based on vertices
- get perspective transform matrix
- warp the perspective getting bird's eye view in a process
Find ROIs of potential characters on every potential license plate
- find contours on the image
- get bounding box of each contour
- discard all small bounding boxes and bounding boxes that don't match character width to height ratio
- discard contours in contours
Recognize characters in given ROIs
- sort potential characters ROIs from left to right
- compare ROI of each potential character with reference characters and get closest matching
- if there's more ROIs than maximum license plate length, delete character with weakest matching
Fill empty characters on license plate
- if there is no maximum number of characters found, fill empty spaces with "?". "?" are filled in proper space based on distance between each character
Check number of characters in first part of license plate
- compute distance between ROIs to determine number of characters in first part of license plate
Check if license plate match polish rules
- change forbidden characters to theirs closest counterpart. For example 2 -> Z in first part of license plate