Reading QR Codes

Every crate at Orchard Vision has to be tracked from the tree to the supermarket. The owner asks if you can use AI to read the crate labels. You could. Or you could print a QR code on each crate, and read it with three lines of code. The best vision engineers change the problem whenever they can.

• Track every crate
• Use AI to read labels?
• Or change the problem

Reading a code

We generate a QR code, holding a crate number, and add the white border that it needs. Then the detector finds and reads it. For a tougher test, we blur the code, and rotate it by twenty five degrees. Run it. Both read perfectly. That's because QR codes were designed to be easy for machines. They have big corner squares for finding the code, and built in error correction, so that even a code with a damaged corner can still be read. So when you control what's being photographed, add markers, and save yourself months of machine learning.

import cv2
import numpy as np

code = cv2.QRCodeEncoder.create().encode("CRATE-4471")
img = cv2.resize(code, (300, 300), interpolation=cv2.INTER_NEAREST)
img = cv2.copyMakeBorder(img, 40, 40, 40, 40, cv2.BORDER_CONSTANT, value=255)

text, points, _ = cv2.QRCodeDetector().detectAndDecode(img)
print("Read:", text)

blurred = cv2.GaussianBlur(img, (9, 9), 0)
turned = cv2.warpAffine(blurred, cv2.getRotationMatrix2D((190, 190), 25, 1), (380, 380), borderValue=255)
print("Rotated and blurred:", cv2.QRCodeDetector().detectAndDecode(turned)[0])

Read: CRATE-4471
Rotated and blurred: CRATE-4471

Change the problem

This is a real mindset. Can't detect a shiny part? Change the light. Can't read a label? Print a code. Can't tell objects apart? Paint a mark on them. Warehouse robots follow QR stickers on the floor, because that's much more reliable than understanding a room. Before you train anything, ask whether you can control the scene.

Summary