Hi! I'm facing a problem of not being able to predict which cluster contains the image i require in a k-means clustering method. I've enclosed the details of the problem.

I'm doing an image segmentation using k-means algorithm to extract just a particular feature from a whole image, for instance a car from the surrounding environment. I've taken three clusters for this. I need to save just the cluster that contains the car, but not able to isolate it since the clusters are generated randomly.

[jgg]

There must be some property which defines a car versus something else; if you don't have that, then aren't you essentially just trying to do image detection (which is much harder)?

[Meera Girijan]

@jgg Based on the color and its intensities in the image each component gets separated.Image detection is what i wanted, but this also gave me a desired output. In the image attached below you can see the car separated from the surroundings in cluster 1.

[Meera Girijan]

@jgg if you could suggest some other efficient method for image detection it would be helpful too. I need to model this separated image.

Answer 1

[Walter Roberson]

You cannot use k-means clustering alone to detect the presence or absence of something. k-means clustering is going to find the given number of clusters without having any interpretation of what the clusters mean. If you feed it an image of a banana instead of an image of a car, it is going to find 3 clusters in the image of the banana.

You should therefore be extracting feature descriptions from portions of the image and run clustering on the feature descriptions -- possibly only on two clusters ("matches" and "does not match") You can then given an interpretation to the clustering based upon knowing what the features imply.

[Meera Girijan]

So, i should do an image detection first and then run the clustering on the detected object?

[Image Analyst]

I agree with Walter. You can't do "image understanding" like is used in CBIR with a simple kmeans analysis of the image itself. You need to derive a bunch of features and then use sophisticated matching algorithms. I suggest you read up on CBIR, but read fast because it's a complicated field and getting much much more so as the algorithms become better and better. It's a rapidly evolving/improving field with a ton of extremely brilliant people working in it. You would be amazed at what the best of them can do now.

[Meera Girijan]

Thank you very much!

[Meera Girijan]

@image analyst Hi!Just wanted to get some more information. Part of my project requires capturing a 2D real time image and extracting a car alone from it without any of the background details. I read up on CBIR and found a lot about image detection. I tried using SURF implementation also. I was getting the interested points, but retrieving the image back seems to be a challenge.The input image for our project will be dynamic and will keep changing every time so I am not able to fix a contour line even if I try to calculate the interested points based on that. I have attached the output i get from kmeans clustering based on colour. The only short coming is that this works only when there's a clear distinguishing color combination . car6 is the actual image and car6sample1 is the output. that is how i want the output to be even when there is no distinct color difference.Could you help me out with this?