Divide image to single objects (coins) for machine learning

Question

I want to automatically divide an image of multiple coins to the single coins, so that afterwards the single coins can be put into a model that classifies the coins (with Tensorflow/Keras).

The input images look something like this

And they should look like this (Sorry that I cannot integrate the images directly as I'm new to StackOverflow).

I want to divide the input images, put the single coins into a classification model, so that I know the single value of the single coins and thereby can identify the value of the first input image.

I already tried an object detection model, but it didn't detect coins (https://towardsdatascience.com/object-detection-with-10-lines-of-code-d6cb4d86f606). As I already know that all objects on the image are coins, I thought that there is maybe an easier way to divide the image?

Thank you in advance.

The proper path is to use an object detection model to define the coin boundaries. As far as I know, there isn't a simpler way. However, there are a number of good object detection libraries that will do this for you (ex: [Tensorflow's object detection API](https://github.com/tensorflow/models/tree/master/research/object_detection). Can you share what you've already tried? — Myles Baker, Sep 05 '18 at 19:31
You want to use the distance transform + watershed method. See for example https://stackoverflow.com/questions/25789278/coffee-beans-separation-algorithm or https://stackoverflow.com/questions/19385264/image-segmentation-of-connected-objects-with-watershed — Cris Luengo, Sep 05 '18 at 22:23

Herr von Wurst · Answer 1 · 2018-09-05T21:13:04.873

I would try color segmentation similar to this tutorial as a first step to separate the coins from the background. Here is my quick try in Python using OpenCV:

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

img = cv2.imread("coins.jpg")

lower = np.array([0,40,5])
upper = np.array([255,255,255])
mask = cv2.inRange(hsv, lower, upper)

cv2.imshow(img)
plt.show()
plt.imshow(mask)

This brings us from the input image

to this mask:

From here using blob analysis and a size filter you should be able to find and separate the unconnected coins. Disconnecting the overlapping areas could be achieved using active contours, or since your goal is to create a training data set, by shifting the coins before taking the picture.

Interesting approach, I will try this out. Thank you! – chris Sep 05 '18 at 20:58 — chris, Sep 05 '18 at 20:58

score 0 · Answer 2 · answered May 29 '20 at 21:49

import cv2
import imutils
import numpy as np
import matplotlib.pyplot as plt

def display(img,count,cmap="gray"):
    f_image = cv2.imread("coins.jpg")
    f, axs = plt.subplots(1,2,figsize=(12,5))
    axs[0].imshow(f_image,cmap="gray")
    axs[1].imshow(img,cmap="gray")
    axs[1].set_title("Total Money Count = {}".format(count))


image = cv2.imread("coins.jpg")
image_blur = cv2.medianBlur(image,25)
image_blur_gray = cv2.cvtColor(image_blur, cv2.COLOR_BGR2GRAY)
image_res ,image_thresh = cv2.threshold(image_blur_gray,240,255,cv2.THRESH_BINARY_INV)
kernel = np.ones((3,3),np.uint8)
opening = cv2.morphologyEx(image_thresh,cv2.MORPH_OPEN,kernel)
dist_transform = cv2.distanceTransform(opening,cv2.DIST_L2,5)
ret, last_image =  cv2.threshold(dist_transform, 0.3*dist_transform.max(),255,0)
last_image = np.uint8(last_image)

cnts = cv2.findContours(last_image.copy(), cv2.RETR_EXTERNAL,
    cv2.CHAIN_APPROX_SIMPLE)
cnts = imutils.grab_contours(cnts)


for (i, c) in enumerate(cnts):
    ((x, y), _) = cv2.minEnclosingCircle(c)
    cv2.putText(image, "#{}".format(i + 1), (int(x) - 45, int(y)+20),
        cv2.FONT_HERSHEY_SIMPLEX, 2, (255, 0, 0), 5)
    cv2.drawContours(image, [c], -1, (0, 255, 0), 2)

display(image,len(cnts))

Divide image to single objects (coins) for machine learning

2 Answers2