Finding by Template

Want to find a logo on a package? Take a picture of the logo, slide it across the image, and see where it fits best. That's template matching. It's simple and fast, and it works perfectly when objects look exactly the same each time, such as parts on a factory line.

• Slide a picture of what you want
• See where it fits best
• Perfect for identical parts

Template matching

We cut one coin out of the image, to use as the template. Match template slides it everywhere, and gives a score at each position. Run it. The best match, of course, is the coin itself, with a score of one. Other coins also score well, since they look similar. But it breaks if an object is rotated, scaled, or lit differently. For that, we need something smarter.

import cv2
import numpy as np
from skimage import data

gray = data.coins()
template = gray[170:220, 75:130]

scores = cv2.matchTemplate(gray, template, cv2.TM_CCOEFF_NORMED)
_, best, _, where = cv2.minMaxLoc(scores)
print("Best match at:", where, " score:", round(best, 2))

ys, xs = np.where(scores > 0.8)
print("Other strong matches:", len(xs))

Best match at: (75, 170)  score: 1.0
Other strong matches: 18

Finding circles

What if you want to find a shape, rather than an exact picture? The Hough transform finds lines and circles. Each edge pixel votes for the circles that could pass through it. Where many votes agree, there's a real circle. Run it. It found the coins, purely from their shape. For round apples on a belt, that's a useful trick, as it copes with some overlap, which our earlier contour method couldn't.

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

gray = data.coins()
blur = cv2.medianBlur(gray, 5)
found = cv2.HoughCircles(blur, cv2.HOUGH_GRADIENT, dp=1.2, minDist=25,
                         param1=120, param2=32, minRadius=15, maxRadius=45)

out = cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB)
for x, y, r in found[0].astype(int):
    cv2.circle(out, (x, y), r, (0, 255, 0), 2)
plt.figure(figsize=(8, 4)); plt.imshow(out); plt.axis("off")
plt.tight_layout(); plt.savefig("circles.png", dpi=110)
print("Circles found:", found.shape[1])

Circles found: 24

Summary