A deep learning based Number Plate Recognition System.
Collecting the data
I developed python webscrapers to crawl through internet and collected around 10,000 images of cars in all positions and angles with their number plate visible to train an accurate ML model without overfitting
Developing the image Recognition software
Yolov2 was used, implemented in python. Trained it to around 5k epox taking almost 3 days to prepare on GTX 1050 graphic hardware.
Developing an unique OCR pipeline
I developed an unique ocr pipeline so the detected image of number plate can be passed through this OCR system and i can detect the letters in the cropped image of the plate. A number of image processing methdologies were used including adaptive thresholding and a different forms of binarization and smoothening.
Technologies used and skills learned were:
- Python3
- Tensorflow
- Web scraping
- YoloV2
- iot
- Raspberri Pi
- OCR
- Tesseract
- Extensive Research on Image Processing