Thresholding

The simplest way to find objects is to pick a brightness. Anything above it is an object, and anything below is background. It's so simple that it sounds like a toy, and yet it's used in factories all over the world, because with good lighting, it just works.

• Pick one brightness
• Above is object, below is background
• Used in factories everywhere

Reading the histogram

A histogram counts how many pixels have each brightness. If your objects are bright and the background's dark, you'll see two hills. The best threshold is the valley between them. In nineteen seventy nine, Nobuyuki Otsu found a way for a computer to pick that valley automatically.

Automatic thresholds

Here we have a photo of coins. Otsu's method finds the threshold for us. But look at the lighting in this photo. It's uneven, so a single threshold fails on one side. Adaptive thresholding picks a different value for each small region. Run it. Compare the two. Otsu loses part of the image, because of the shadow. Adaptive handles the changing light, though it does add some noise. There's always a trade-off.

import cv2
import matplotlib.pyplot as plt
from skimage import data

gray = data.coins()

value, otsu = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
adaptive = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
                                 cv2.THRESH_BINARY, 51, 2)

fig, ax = plt.subplots(1, 3, figsize=(10, 3))
for a, im, n in zip(ax, [gray, otsu, adaptive], ["Original", "Otsu", "Adaptive"]):
    a.imshow(im, cmap="gray"); a.set_title(n); a.axis("off")
plt.tight_layout(); plt.savefig("threshold.png", dpi=110)
print("Otsu chose the value:", value)

Otsu chose the value: 107.0

Summary