Coffee beans separation algorithm

Below is presented an approach to find the center of each bean. Analyzing the central position of segmented objects in frames in different, but sequential, time it is possible to track them. Keeping visual profiles or analyzing its path can increase the accuracy of the tracking algorithm in situations that an object cross the other or there are some overlap.

I used Marvin Image Processing Framework and Java.

Finding the center approach

I used three basic algorithms: threshold, morphological erosion and floodfill segmentation. The first step is the threshold for removing the background, as shown below.

Thresholding

The next step is the application of morphological erosion in order to separate the beans. In the case of a small kernel matrix I can separate the small beans but keep the bigger ones together, as shown below. Filtering using the mass (number of pixels) of each independent segment it is possible to select just the smaller ones, as shown below.

Erosion small kernel

Using a big kernel matrix I can separate the bigger ones and the small ones disappear, as shown below.

enter image description here

Combining the two results - removing center points that are too near and probably from the same bean - I got the result presented below.

enter image description here

Even not having the real segment of each bean, using the center positions it is possible to count and track them. The centers can also be used to find out each bean segment.

Source code

The source code is in Java, but the image processing algorithms employed in the solution are provided by the most frameworks.


EDIT: I edited the source code in order to save the images of each step. The source code can be optimized removing these debug steps and creating methods to reuse code. Some objets and lists were created just to demonstrate theses steps and can be removed too.
import static marvin.MarvinPluginCollection.floodfillSegmentation;
import static marvin.MarvinPluginCollection.thresholding;
import marvin.image.MarvinColorModelConverter;
import marvin.image.MarvinImage;
import marvin.image.MarvinSegment;
import marvin.io.MarvinImageIO;
import marvin.math.MarvinMath;
import marvin.plugin.MarvinImagePlugin;
import marvin.util.MarvinPluginLoader;

public class CoffeeBeansSeparation {

    private MarvinImagePlugin erosion = MarvinPluginLoader.loadImagePlugin("org.marvinproject.image.morphological.erosion.jar");

    public CoffeeBeansSeparation(){

        // 1. Load Image 
        MarvinImage image = MarvinImageIO.loadImage("./res/coffee.png");
        MarvinImage result = image.clone();

        // 2. Threshold
        thresholding(image, 30);

        MarvinImageIO.saveImage(image, "./res/coffee_threshold.png");

        // 3. Segment using erosion and floodfill (kernel size == 8)
        List<MarvinSegment> listSegments = new ArrayList<MarvinSegment>();
        List<MarvinSegment> listSegmentsTmp = new ArrayList<MarvinSegment>();
        MarvinImage binImage = MarvinColorModelConverter.rgbToBinary(image, 127);

        erosion.setAttribute("matrix", MarvinMath.getTrueMatrix(8, 8));
        erosion.process(binImage.clone(), binImage);

        MarvinImageIO.saveImage(binImage, "./res/coffee_bin_8.png");
        MarvinImage binImageRGB = MarvinColorModelConverter.binaryToRgb(binImage);
        MarvinSegment[] segments =  floodfillSegmentation(binImageRGB);

        // 4. Just consider the smaller segments
        for(MarvinSegment s:segments){
            if(s.mass < 300){   
                listSegments.add(s);
            }
        }

        showSegments(listSegments, binImageRGB);
        MarvinImageIO.saveImage(binImageRGB, "./res/coffee_center_8.png");

        // 5. Segment using erosion and floodfill (kernel size == 18)
        listSegments = new ArrayList<MarvinSegment>();
        binImage = MarvinColorModelConverter.rgbToBinary(image, 127);

        erosion.setAttribute("matrix", MarvinMath.getTrueMatrix(18, 18));
        erosion.process(binImage.clone(), binImage);

        MarvinImageIO.saveImage(binImage, "./res/coffee_bin_8.png");
        binImageRGB = MarvinColorModelConverter.binaryToRgb(binImage);
        segments =  floodfillSegmentation(binImageRGB);

        for(MarvinSegment s:segments){
            listSegments.add(s);
            listSegmentsTmp.add(s);
        }

        showSegments(listSegmentsTmp, binImageRGB);
        MarvinImageIO.saveImage(binImageRGB, "./res/coffee_center_18.png");

        // 6. Remove segments that are too near.
        MarvinSegment.segmentMinDistance(listSegments, 10);

        // 7. Show Result
        showSegments(listSegments, result);
        MarvinImageIO.saveImage(result, "./res/coffee_result.png");
    }

    private void showSegments(List<MarvinSegment> segments, MarvinImage image){
        for(MarvinSegment s:segments){
            image.fillRect((s.x1+s.x2)/2, (s.y1+s.y2)/2, 5, 5, Color.red);
        }
    }

    public static void main(String[] args) {
        new CoffeeBeansSeparation();
    }
}

This approach is a spin-off from mmgp's answer that explains in detail how the watershed algorithm works. Therefore, if you need some explanation on what the code does, please check his answer.

The code can be played with in order to improve the rate of detection. Here it is:

import sys
import cv2
import numpy
from scipy.ndimage import label

def segment_on_dt(a, img):
    border = cv2.dilate(img, None, iterations=3)
    border = border - cv2.erode(border, None)
    cv2.imwrite("border.png", border)

    dt = cv2.distanceTransform(img, 2, 5)    
    dt = ((dt - dt.min()) / (dt.max() - dt.min()) * 255).astype(numpy.uint8)
    _, dt = cv2.threshold(dt, 135, 255, cv2.THRESH_BINARY)
    cv2.imwrite("dt_thres.png", dt)    

border (left), dt (right):

enter image description hereenter image description here

    lbl, ncc = label(dt)
    lbl = lbl * (255/ncc)      
    # Completing the markers now. 
    lbl[border == 255] = 255

    lbl = lbl.astype(numpy.int32)
    cv2.imwrite("label.png", lbl)

lbl:

enter image description here

    cv2.watershed(a, lbl)

    lbl[lbl == -1] = 0
    lbl = lbl.astype(numpy.uint8)
    return 255 - lbl

# Application entry point
img = cv2.imread("beans.png")
if img == None:
    print("!!! Failed to open input image")
    sys.exit(0)

# Pre-processing.
img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)    
_, img_bin = cv2.threshold(img_gray, 128, 255, cv2.THRESH_OTSU | cv2.THRESH_BINARY_INV)
cv2.imwrite("img_bin.png", img_bin)

img_bin = cv2.morphologyEx(img_bin, cv2.MORPH_OPEN, numpy.ones((3, 3), dtype=int))
cv2.imwrite("img_bin_morphoEx.png", img_bin)

img_bin (left) before and after (right) a morphology operation:

enter image description hereenter image description here

result = segment_on_dt(img, img_bin)
cv2.imwrite("result.png", result)

result[result != 255] = 0
result = cv2.dilate(result, None)
img[result == 255] = (0, 0, 255)
cv2.imwrite("output.png", img)

result (left) of watershed segmentation, followed by output (right):

enter image description hereenter image description here


Here is some code (in Python) that will give you a baseline. Count the number of black pixels and divide into the area accounting by how many circles of average size can be packed into a square of your size. The has the virtue of being the simplest possible thing you can do.

If a given method is not on average more accurate than this, then you need a better method. BTW I'm getting around 85% accuracy, so your 95% is not out of the question.

import Image

im = Image.open('ex2a.gif').convert('RGB')
(h,w) = im.size
print h,w
num_pixels = h*w
print num_pixels
black_pixels = 0
for i in range(h):
    for j in range(w):
        q = im.getpixel((i,j)) 
        if q[0]<10 and q[1]<10 and q[2]<10:
            black_pixels = black_pixels + 1
            im.putpixel((i,j),(255,0,0))
r = 15
unpackable = (h/(2*r))*(w/(2*r))*((2*r)**2 - 3.14*r**2)
print 'unpackable:',unpackable
print 'num beans:',round((num_pixels-2*unpackable)/750.0)
im.save('qq.jpg')

There are some elegant answers, but I thought of sharing what I tried because it is bit different to other approaches.

After thresholding and finding the distance transform, I propagate the local maxima of the distance-transformed image. By adjusting the extent of maxima propagation I segment the distance transformed image, then filter these segments by their area, rejecting smaller segments.

This way I can achieve a reasonably good segmentation of the given image, though it does not clearly define the boundaries. For the given image I get a segment count of 42 using the parameter values that I use in the Matlab code to control the extent of maxima propagation and the area threshold.

Results:

enter image description here

enter image description here

Here's the Matlab code:

clear all;
close all;

im = imread('ex2a.gif');
% threshold: coffee beans are black
bw = im2bw(im, graythresh(im));
% distance transform
di = bwdist(bw);
% mask for coffee beans
mask = double(1-bw);

% propagate the local maxima. depending on the extent of propagation, this
% will transform finer distance image to coarser segments 
se = ones(3);   % 8-neighbors
% this controls the extent of propagation. it's some fraction of the max
% distance of the distance transformed image (50% here)
mx = ceil(max(di(:))*.5);
peaks = di;
for r = 1:mx
    peaks = imdilate(peaks, se);
    peaks = peaks.*mask;
end

% how many different segments/levels we have in the final image
lvls = unique(peaks(:));
lvls(1) = []; % remove first, which is 0 that corresponds to background
% impose a min area constraint for segments. we can adjust this threshold
areaTh = pi*mx*mx*.7;
% number of segments after thresholding by area
nseg = 0;

% construct the final segmented image after thresholding segments by area
z = ones(size(bw));
lblid = 10;  % label id of a segment
for r = 1:length(lvls)
    lvl = peaks == lvls(r); % pixels having a certain value(level)
    props = regionprops(lvl, 'Area', 'PixelIdxList'); % get the area and the pixels
    % threshold area
    area = [props.Area];
    abw = area > areaTh;
    % take the count that passes the imposed area threshold
    nseg = nseg + sum(abw);
    % mark the segments that pass the imposed area threshold with a unique
    % id
    for i = 1:length(abw)
        if (1 == abw(i))
            idx = props(i).PixelIdxList;
            z(idx) = lblid; % assign id to the pixels
            lblid = lblid + 1; % increment id
        end
    end
end

figure,
subplot(1, 2, 1), imshow(di, []), title('distance transformed')
subplot(1, 2, 2), imshow(peaks, []), title('after propagating maxima'), colormap(jet)
figure,
subplot(1, 2, 1), imshow(label2rgb(z)), title('segmented')
subplot(1, 2, 2), imshow(im), title('original')