Detecting if image matches another

[kainev]

Hi,

I have a program that continually takes screenshots of a Nintendo 64 Game from a capture card, it needs to compare the screenshot and determine if it matches another given screenshot.

My original plan was to use the compare_ssim method from the skimage.measurepackage to determine the structural similarity, and then if the ssim percentage was high enough (i.e 95%) say it's match. This works well in theory, however the quality of capture cards vary and ssim gives poor results in practice.

My SSIM test code:

from skimage.measure import compare_ssim
import cv2

screenshot_img = cv2.cvtColor(cv2.imread("screenshot.png"), cv2.COLOR_BGR2GRAY)
comparison_img = cv2.cvtColor(cv2.imread("comparison.png"), cv2.COLOR_BGR2GRAY)

ssim = calc_ssim(screenshot_img, comparison_img)
print(ssim)

Here are two examples of what I'm trying to determine match.

[Image: zilz5w.png]

[Image: zwf1ph.jpg]

See how the top one is lower quality, the ssim only rates these as a 30% match, however I need my program to recognise these are images of the same thing with certainty.

What else can I do to reliably determine whether two images such as the ones posted above are the same/very similar? I'm open to a new method entirely. It just has to be somewhat quick as my application has to process at 29.97fps, and around half the frame is already spent taking the screenshot.

Thanks!

[wavic]

Just compare md5 or sha1 sum of the two files.

[kainev]

hey, should have mentioned, as the program is taking screenshots quickly and doing processing on them, I never actually save them to disk. Only the comparison image is saved to disc. The code posted was just test code as stated to test the idea.

EDIT:

I may have found a solution, I'm using openCVs matchTemplate function, and passing in a cropped version of the logo posted in my original thread as the template. Something like this:

res = cv2.matchTemplate(image, template_image, cv2.TM_SQDIFF_NORMED)
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)

if min_val < 0.1:
    print("Match found")

If anyone has another suggestion or an improvement please let me know, this is still just early testing.