Reading Text with OCR

Every crate at Orchard Vision has a printed label, with a batch number and a date. Staff type these in by hand. Mistakes cost money when food has to be recalled. Reading text in images is called optical character recognition, and it's one of the oldest dreams in the field.

• Every crate has a label
• Typed in by hand
• Mistakes cost money

How modern OCR works

Modern OCR has two stages. First, a detector finds regions of text. Then a recogniser reads each region, often using a convolutional network followed by a sequence model. Good results depend on clean input. Straighten the page, boost the contrast, and keep the text large enough. All of the classic tools that you learned earlier help here.

Reading a page

Here's a real engine, reading a scanned page. It gives each line of text, its position, and a confidence score. Run it. It's reading a page with old printing and uneven light, and it does well. The confidence score matters, because you can send low confidence lines to a person.

import cv2
from rapidocr_onnxruntime import RapidOCR
from skimage import data

engine = RapidOCR()
image = data.page()

result, _ = engine(cv2.cvtColor(image, cv2.COLOR_GRAY2BGR))

print("Lines read:", len(result))
for box, text, score in result[:4]:
    print(f"{float(score):.2f}  {text[:60]}")

Lines read: 5
0.94  Region-basedsegmentation
0.92  Let us first determine markers of the coins and the
0.89  background. These markers are pixels that we can label
0.91  unambiguously as either object or background. Here,

Reading a crate label

Now AppleEye reads a label. We draw one, read it, then use regular expressions to pull out the batch and date. Run it. Clean, structured data, straight from an image. Combining OCR with strict patterns is powerful, because the pattern catches many reading errors. A batch number that doesn't match gets flagged for a human.

import cv2
import numpy as np
from rapidocr_onnxruntime import RapidOCR

img = np.full((180, 520, 3), 245, np.uint8)
cv2.putText(img, "BATCH 2291", (20, 70), cv2.FONT_HERSHEY_SIMPLEX, 1.4, (20, 20, 20), 3)
cv2.putText(img, "BEST BY 12/2026", (20, 140), cv2.FONT_HERSHEY_SIMPLEX, 1.2, (20, 20, 20), 3)

result, _ = RapidOCR()(img)
text = " ".join(line[1] for line in result)
print("Read:", text)

import re
print("Batch:", re.findall(r"BATCH\s*(\d+)", text))
print("Date: ", re.findall(r"\d{2}/\d{4}", text))

Read: BATCH  2291 BEST BY 12/2026
Batch: ['2291']
Date:  ['12/2026']

Summary