run notebook EDA.ipynb

10% of train data moved to validation

1. classes frequencies in train and test are different
2. some symbols are unlabelled: check pictures with less than 7 objects
3. Objects are usually have similar size - easier to learn anchors.
4. Most pictures are aligned (sides are parallel to screen sides). But some are not
5. 90% of photos has resolution less than 630 x 324. Preprocessing with resize to 640 looks optimal
6. objects are distributed on

run train_YOLO_on_arabic_plates.ipynb

same notebook {'metrics/precision(B)': 0.9208474724456436, 'metrics/recall(B)': 0.9656120747252827, 'metrics/mAP50(B)': 0.9925618014464169, 'metrics/mAP50-95(B)': 0.7770107902541031, 'fitness': 0.7985658913733344}

check and correct data. Some mislabelled cases found. It lead to FalsePositive results and affects classification Precision. Generate synthetic data Get rid of low res photos

Try other nets: FasterRCNN, DETR, Try detectors pretrained on OCR. For example from PaddleOCR.