TrOCR (OCR)

TrOCR is a transformer-based model for text recognition, otherwise known as Optical Character Recognition (OCR).

TrOCR works best on focused, single-line printed text.

Be sure to use with cropped images since unlike some other OCR models, TrOCR will not perform well on uncropped or multi-line text.

Let's try running TrOCR on this image:

Serial number

Note

TROCR model is only supported in inference Python package and inference server deployed locally (excluding Roboflow Hosted Platform).

To run the example, start inference server locally:

inference server start

Make sure you have inference-cli installed - if that's not the case run:

pip install inference-cli

import os
from inference_sdk import InferenceHTTPClient

CLIENT = InferenceHTTPClient(api_url="http://127.0.0.1:9001")

result = CLIENT.ocr_image(inference_input="./serial_number.png", model="trocr")  # single image request
print(result)