opencv pattern matching not works

Question

I need find all images within image, for this idea I have found great solution:

import cv2
import numpy as np

img_rgb = cv2.imread('source.png')
img_gray = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
template = cv2.imread('block.png', 0)
w, h = template.shape[::-1]

res = cv2.matchTemplate(img_gray, template, cv2.TM_CCOEFF)

threshold = 0.8
loc = np.where(res >= threshold)

for pt in zip(*loc[::-1]):
    cv2.rectangle(img_rgb, pt, (pt[0] + w, pt[1] + h), (0, 0, 255), 2)

# cv2.imwrite('res.png', img_rgb)
cv2.imshow('output', img_rgb)
cv2.waitKey(0)

Source data:

https://i.sstatic.net/cE5bM.png (source)

https://i.sstatic.net/BgzAA.png (template)

I tried to use this code but failed.

What I see now:

What I expected to get:

What's wrong? I am using python 3.5 and opencv 3.3.0.10

PS: very interesting thing that another solution works perfect but finds only 1 match (best one)

Answer 1

I am definitely no expert on OpenCV and it's various template matching methods (though coincidentally I had started to play around with it).

However, a couple of things in your example stand out.

You use the cv2.TM_CCOEFF method which gives results that are universally way above the 0.8 threshold. So everywhere in the image matches giving a massive red rectangle blob. If you want to use this method try cv2.TM_CCOEFF_NORMED to normalise the results to below 1.

But my best 10 minute attempt was using;

method = cv2.TM_CCORR_NORMED

and setting

threshold = 0.512

which gave;

This is fairly unsatisfactory though because the threshold had to be 'tuned' fairly precisely to remove most of the mismatches. There is undoubtedly a better way to get a more reliable stand-out match.