I think I have understood the DBScan algorithm for 2D data points. We can consider the example in scikit-learn. They generate a set of data points:

from sklearn.datasets import make_blobs
from sklearn.preprocessing import StandardScaler

centers = [[1, 1], [-1, -1], [1, -1]]
X, labels_true = make_blobs(
    n_samples=750, centers=centers, cluster_std=0.4, random_state=0

X = StandardScaler().fit_transform(X)

The X and Y data points are:

[[ 0.49426097  1.45106697]
 [-1.42808099 -0.83706377]
 [ 0.33855918  1.03875871]
 [-0.05713876 -0.90926105]
 [-1.16939407  0.03959692]
 [ 0.26322951 -0.92649949]]

The parameters to be chosen for the DBScan algorithm are two: neighborhood radius ϵ and minimum number of samples. It works in this way: the algorithm starts from one random sample and calculates how many other samples fall within its neighborhood radius ϵ and generates a cluster if in the neighborhood radius there are at least the minimum number of samples.

So for the example above:

from sklearn.cluster import DBSCAN

db = DBSCAN(eps=0.3, min_samples=10).fit(X)
labels = db.labels_

import numpy as np

labels = np.unique(labels)

Out[7]: array([-1,  0,  1,  2], dtype=int64)

It has found 3 clusters and some noise points.

Now, if I want to apply this in order to cluster groups in images, how can I do ?

I found this example here https://www.youtube.com/watch?v=wyk_vkL2os8. I tried to reproduce it using one image example that I found here in StackOverflow.

from sklearn.cluster import DBSCAN
from matplotlib.image import imread
import numpy as np
import matplotlib.pyplot as plt

#load image
image = imread('pZKmf.png')#take the image and convert into pixels using imread

print(image.shape)#(217, 386, 4) 


The original image:

enter image description here

#convert it into two dimensional
#Flatten the image to create a 2D array of pixels
# you rescale in 2D array in order to have (features,samples)
X = image.reshape(-1,4)

print(X.shape) #(83762, 4)

#Apply the DBSCAN algorithm

dbscan = DBSCAN(eps=0.01, min_samples=500)

labels = dbscan.fit(X)

print(labels.shape) #(83762,)

unique_labels = np.unique(labels) 
#it corresponds to the number of clusters
#array([-1,  0,  1,  2], dtype=int64)
#It has found 3 clusters

segmented_img = labels.reshape(image.shape[:2])

print(segmented_img.shape) #(217, 386)


Here it is the segmented image:

enter image description here

In this example, it considers pixel intensities as features and we have 4 images which are the 4 samples.

I can not figure out which are the samples to cluster... I would consider pixels as samples to cluster. How can I understand how does clustering work in this case ?


2 Answers 2


One important step in clustering images is how the image is encoded. One useful way to encode an image is using vector embeddings, where the choosing numerical values have semantic meaning. The closer two images are in the vector embedding space the more meaning they share, thus the clustering captures the meaning of the images. There are many models available to create image embeddings, Hugging Face is one example.



[2] https://emad-ezzeldin4.medium.com/debugging-computer-vision-image-classification-why-is-your-model-failing-in-production-11976e5311f2?source=friends_link&sk=5c49481d65179689093b4b4fb9b8e231


This [1] is an actual example of images clustered well using vector embeddings of Vgg16 image classifier as features ( as opposed to pixel intensity in your post). It was clustered with KNN though and not image segmentation per your example. DBSCAN is probably powering engines like LIME (local interpretable model-agnostic explanations) to find the highest density of high value features within the image [2]. To simplify , we can think of it as a convolution window scanning the image for important features used by the classifier. Same principles are applicable in image segmentation probably.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.