Find positions of objects on the noisy image

Question

I have a bunch of images and I need to determine positions of crosses for further transformation of the image and the alignment procedure. The problem is that images are quite noisy and I'm new to all these things of computer vision. Generally, I'm trying to solve the task via opencv and python. I have tried several approaches described in the tutorial of opencv library but I did not get the appropriate result.

Consider: I need to determine the exact positions of centers of the crosses (which I can do with about pixel accuracy by hand). The best result I have obtained via findContours function. I have adopted code from the tutorial and I got:

import numpy as np
import cv2 as cv
import matplotlib.pyplot as plt
import random

random.seed(42)

img = cv.imread("sample.png")
img_gray = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
img_gray = cv.blur(img_gray, (3,3))

threshold = 150

dst = cv.Canny(img_gray, threshold, threshold * 2)
_, contours, hierarchy = cv.findContours(dst, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)

result = np.zeros((dst.shape[0], dst.shape[1], 3), dtype=np.uint8)
for i in range(len(contours)):
    color = (random.randint(0, 256), random.randint(0, 256), random.randint(0, 256))
    cv.drawContours(result, contours, i, color, 2, cv.LINE_8, hierarchy, 0)

cv.imwrite("result.png", result)

fig, ax = plt.subplots()
fig.set_size_inches(10, 10);
ax.imshow(result, interpolation='none', cmap='gray');

which results in: Now I'm confused with the following steps. How can I define which contour is cross and which is not? What to do with crosses consisting of multiple contours?

Any help is really appreated!

Answer 1

A simple way on which you can determine what is a cross and what isn't is by making a bouning box x,y,w,h = cv2.boundingRect(cnt) over each contour and selecting those that have h (height) and w (weight) bigger than treshold you provide. If you observe the noises on the image arent as big as the crosses.

I have also made an example on how I would try to tackle such a task. You can try denoising the image by performing histogram equalization followed by thresholding with OTSU threshold and performing an opening to the threshold (erosion followed by dilation). Then you can filter out crosses with height and weight of the contour and then calculate the middle points of every bounding box of the contours that is in the mentioned criteria. Hope it helps a bit. Cheers!

Example:

import cv2
import numpy as np

img = cv2.imread('croses.png')
gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
equ = cv2.equalizeHist(gray)
_, thresh = cv2.threshold(equ,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
kernel = np.ones((2,2),np.uint8)
opening = cv2.morphologyEx(thresh,cv2.MORPH_OPEN,kernel, iterations = 2)
_, contours, hierarchy = cv2.findContours(opening,cv2.RETR_TREE,cv2.CHAIN_APPROX_NONE)

for cnt in contours:
    x,y,w,h = cv2.boundingRect(cnt)
    if w > 40 and h > 40:
        cv2.rectangle(img,(x,y),(x+w,y+h),(0,255,0),2)
        cv2.circle(img,(int(x+(w/2)), int(y+(h/2))),3,(0,0,255),-1)

cv2.imshow('img', img)
cv2.waitKey(0)
cv2.destroyAllWindows()

Result: