University of Sciences
Subject: Introduction to Computer Vision
Final project: Automatic License Plate Recognition (ALPR)
We developed 2 versions:
- Version 1: Process for car plate in Europe. Almost based on handcraft algorithm to segment region in image and then extract license plate region. Also we use SVM classifer to class image as Plate or Not. Finally, we use Tesseract library to recognize plate numbers. But I find Tesseract didn't work well with recognizing each single plate character/number, Tesseract just works well with recognizing semantic sentences. Because it use Long Short-Term Memory (LSTM).
- Version 2: Process for motor plate in Viet Nam. We used transfer learning methods with pretrained model from https://thigiacmaytinh.com/modelcascade-da-huan-luyen/. Particularly, we used Haar-Cascade Classifier rely on Haar features and use it for detecting license plate region.
More detail, please check it at our proposal and report document. Those 2 versions don't give the best perfomance comparing with state-of-the-art project in related fields. They are just acceptable results.
- 1612174 Phùng Tiên Hao [email protected]
- 1612269 Võ Quôc Huy [email protected]
- 1612272 Trân Nhât Huy [email protected]
There are proposal and report file to have you can refer them and get more intuitive. Moreover, We also have presentation file. These document was written in Vietnamese.
Handcraft algorithm work kind of well and it detect most correct plate region. But in some difficult case such as noise background, scale (too big/small), light (dazzing or too dark) our system is not able to detect plate.
Haarcascade classifier is actually good because it detects correctly most cases which especially work even in any scale. However, the bounding box was generated by Haarcascade which are quite big comparing with Handcraft algorithm. In general, its perfomance still better than handcraft algorithm. Now, Haarcascade is not as strong as YOLO. We intend to use YOLO with moto dataset but at last we dont have enough time to handle it. So we just use pretrained Haar-Cascade model.