YOLO26 and EasyOCR: number plate recognition tutorial

This tutorial builds an automatic number plate recognition pipeline with Ultralytics Platform, YOLO26 nano and EasyOCR. The presenter starts with a labeled dataset, configures 50 training epochs and selects an A100 for cloud training. The platform also has a local training option that downloads the dataset from the cloud.

The walkthrough covers precision and recall curves, image predictions and deployment region selection. The presenter reports about 99% mAP at an IoU threshold of 0.50 for this training run. That result measures plate detection on the demonstrated dataset, not the accuracy of the complete text recognition pipeline.

The Python script requires a deployment URL and API key. It reads images with OpenCV, encodes them as JPEG bytes and sends them to the cloud detector. After receiving bounding boxes, it pads the plate crops, resizes and thresholds them, then uses EasyOCR to extract text. It saves annotated images in a runs folder.

The final test uses 10 car images and includes OCR errors, such as extra digits. The presenter mentions PaddleOCR and suggests Gemma 4 as a model to run locally for OCR, but does not test either alternative. The demonstrated detector depends on cloud requests; this is not an offline setup.