Writting robust (size invariant) circle detection (Watershed)

edit

I want to detect circles with different sizes. The use case is to detect coins on an image and to extract them solely. Quick Summary so far: I use the watershed algorithm but I have probably a problem with threshold. It didn't detect the brighter circles.

It's probably not ideal but I already have described it in a stackoverflow thread and I would be really happy if I could get any more advices: https://stackoverflow.com/questions/5...

I used this watershed algorithm: https://docs.opencv.org/3.4/d3/db4/tu...

Edit: Example image of my problem.

Edit 2:I tried the radial symmetry transform by kbarni. I used [this] (https://github.com/ceilab/frst_python) implementation but unfortunately I only get a just black output. I have already tried to modify every parameter but it didn't help.

'''
Implementation of fast radial symmetry transform in pure Python using OpenCV and numpy.

Adapted from:
https://github.com/Xonxt/frst

Which is itself adapted from:
Loy, G., & Zelinsky, A. (2002). A fast radial symmetry transform for detecting points of interest. Computer         
Vision, ECCV 2002.
'''

import cv2
import numpy as np


def gradx(img):
    img = img.astype('int')
    rows, cols = img.shape
    # Use hstack to add back in the columns that were dropped as zeros
    return np.hstack((np.zeros((rows, 1)), (img[:, 2:] - img[:, :-2]) / 2.0, np.zeros((rows, 1))))


def grady(img):
    img = img.astype('int')
    rows, cols = img.shape
    # Use vstack to add back the rows that were dropped as zeros
    return np.vstack((np.zeros((1, cols)), (img[2:, :] - img[:-2, :]) / 2.0, np.zeros((1, cols))))


# Performs fast radial symmetry transform
# img: input image, grayscale
# radii: integer value for radius size in pixels (n in the original paper); also used to size gaussian kernel
# alpha: Strictness of symmetry transform (higher=more strict; 2 is good place to start)
# beta: gradient threshold parameter, float in [0,1]
# stdFactor: Standard deviation factor for gaussian kernel
# mode: BRIGHT, DARK, or BOTH
def frst(img, radii, alpha, beta, stdFactor, mode='BOTH'):
    mode = mode.upper()
    assert mode in ['BRIGHT', 'DARK', 'BOTH']
    dark = (mode == 'DARK' or mode == 'BOTH')
    bright = (mode == 'BRIGHT' or mode == 'BOTH')

    workingDims = tuple((e + 2 * radii) for e in img.shape)

    # Set up output and M and O working matrices
    output = np.zeros(img.shape, np.uint8)
    O_n = np.zeros(workingDims, np.int16)
    M_n = np.zeros(workingDims, np.int16)

    # Calculate gradients
    gx = gradx(img)
    gy = grady(img)

    # Find gradient vector magnitude
    gnorms = np.sqrt(np.add(np.multiply(gx, gx), np.multiply(gy, gy)))

    # Use beta to set threshold - speeds up transform significantly
    gthresh = np.amax(gnorms) * beta

    # Find x/y distance to affected pixels
    gpx = np.multiply(np.divide(gx, gnorms, out=np.zeros(gx.shape), where=gnorms != 0),         
    radii).round().astype(int);
    gpy = np.multiply(np.divide(gy, gnorms, out=np.zeros(gy.shape), where=gnorms != 0),     
    radii).round().astype(int);

    # Iterate over all pixels (w/ gradient above threshold)
    for coords, gnorm in np ...