The project has three components:
- Vehilce Detection using YOLO
- License Plate segmentation, done using WPOD-NET (currently under review due to not so encouraging results)
- OCR from License Plates: Done using KNN, Tessaract and CNN to compare the accuracy of the methods.
The project is inspired from https://arxiv.org/abs/1312.6082 , developing CNN based Multi Digit Recognition which performed very well even on the Hardest CAPTCHA and from https://link.springer.com/chapter/10.1007%2F978-3-030-01258-8_36#enumeration , which also contributed greately towards structuring this project. src code is from their repo https://github.com/sergiomsilva/alpr-unconstrained
Another interseting paper I came across while reading through the texts to get better insight on the topic was https://web.stanford.edu/class/cs231m/projects/final-report-yang-pu.pdf that employed the CNN and used the model to build a mobile application for Multi Digit Recognition.
Integration of the modules is ongoing ...