ANPL

There is three main step for detecting persian lisence plate:

• train yolov7 with persian license plate dataset for object detection
• crop and pre processing object detected from yolo
• reading lisence plate

yolov7

in this step we need to train last layers of pre trained yolov7 because coco dataset that yolo trained with dont exist of lisence plates and for better performance we use persian lisence plate dataset.

unfortunatly i didnt found good and enought dataset so i merg 2 dataset together and make a data set with 1000 picture of lisence plates in persian(because roboflow free account limit is 1000 image).problem here is that label of each dataset is diffren so we should rename the labels and make simillar label for 2 datasets we merged for this part i used the code bellow for changing label of annotation files:

from glob import glob
import pandas as pd
from pathlib import Path
for j in range (313):
    path = pd.DataFrame(glob('*'))
    path.columns=['location']
    z=path['location'].iloc[2*j+1]
    a=[]
    with open(z, "r") as f , open(fr'E:/dataset-car-plate/Vehicle Plates/Vehicle Plates/newlabels/{z}','w') as f2:
        a = f.read()
        x= a.replace("vehicle plate", "car")
        f2.write(x)
        f.closed
        f2.closed

now we have datasets for train yolov7. code for training available in objectdetection.ipynb.

OCR

after training yolov7 we jump into step three and train a model for reading the lisence plate detected with yolov7 in this step however you train better model the result we be better after training we save model.

now we have every model we need for reading persian lisence plate

ANPL

we can detect and reading lisence plate from images video or webcam.

input image(or video) first should feed into object detection model we train and saved befor then ddetected isence plate should preprocess and crop and reshape in right size and format for feeding into ocr model after that we feed the preprocessed image into ocr madel we saved before and read the lisence plate in the image(or video)

code for this part available in objectdetection.ipynb.