Ask Your Question

How to improve the accuracy of logo matching?

asked 2018-07-12 05:43:49 -0500

I want to check which logo an image contains. There are 2 logos. In fact, the image contains the first logo but the score of the second logo using sift is higher.

I used BFMatcher to get good match and I also calculated the number of inliers. Both values indicate the second logo is more likely to present.

Are there anythings wrong with my code? How to improve the accuracy?

Thank you very much.

import cv2
import requests
import numpy as np

def get_array(url):
    logo1 = requests.get(url).content
    arr = np.asarray(bytearray(logo1), dtype=np.uint8)
    logo1_array = cv2.imdecode(arr, -1)
    return logo1_array

def get_good_match_num(image_array, logo_array):

    sift = cv2.xfeatures2d.SIFT_create()
    bf = cv2.BFMatcher()

    img_kp, img_des = sift.detectAndCompute(image_array, None)
    logo_kp, logo_des = sift.detectAndCompute(logo_array, None)  

    matches = bf.knnMatch(img_des, logo_des, k=2)

    good = []

    for p in matches:
            m, n = p
            if m.distance < 0.7*n.distance:

    if len(good) > 10:
        src_pts = []
        dst_pts = []
        for m in good:
                a, b = img_kp[m.queryIdx].pt, logo_kp[m.trainIdx].pt

        src_pts = np.float32(np.array(src_pts)).reshape(-1,1,2)
        dst_pts = np.float32(np.array(dst_pts)).reshape(-1,1,2)

        M, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC,5.0)
        num = sum(sum(mask))
        num = 0
    return len(good), num

logo1_array = get_array(r'')
logo2_array = get_array(r'')
image_array = get_array(r'')

good1, num1 = get_good_match_num(image_array, logo1_array) #good1 = 16, num1 = 0
good2, num2 = get_good_match_num(image_array, logo2_array) #good2 = 69, num1 = 51
edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2018-07-12 05:57:18 -0500

berak gravatar image

updated 2018-07-12 06:07:04 -0500

"I want to check which logo an image contains. " -- you can't do it this way.

feature matching will give you a homography Matrix, IF the scene contains the part you're looking for, but you cannot make it say: no, it doesn't.

(you can only solve the "where", not "what" question like this)

it's a common misconception ;(

edit flag offensive delete link more
Login/Signup to Answer

Question Tools

1 follower


Asked: 2018-07-12 05:43:49 -0500

Seen: 45 times

Last updated: Jul 12