Well there are many options, for instance if you have another data set with labelled images having same category of objects you can use transfer learning approach to classify the unlabeled data set. Another approach is Topic Modeling. Classical unsupervised approaches include clustering and you can use more advanced clustering techniques such as migrating means clustering based classification.

Thankyou Sander Ali Khowaja for your quick response,

I have no pre/another data set available and for the inital data gathering could only get us some thousands of images.

I have the images in the format of tiff files and I want my model to learn to classify automatically the samples lets say good or bad class.

Any help will be really appreciated, If you have any algorthim in mind that I shoudl look into as a starting point. Thank you

and using what measures you want to evaluate the good or bad such as PSNR, SSIM or any other. Any way the point is you can used any of the above by defining an objective function or designing a criteria to make the network learn about good or bad image.

you can refer to the following links for further learning.

"http://ufldl.stanford.edu/tutorial/selftaughtlearning/SelfTaughtLearning/ "

"CUNet: A Compact Unsupervised Network for Image Classification "

" Unsupervised Image Ranking "

I hope this helps

Dear Manav Madan

As you mentioned, the unsupervised methods can be useful for your purpose. k-means clustering is a proper method for your approach. However, number of K is challenging. In this regard, you can estimate 'K'(number of clusters) by means of Hierarchical clustering. As the result, a hybrid method(first, estimate number of cluster using Hierarchical clustering, and second , final clustering using K-means) is a applicable approach for your problem.

the following toolbox in Matlab can be used as sample code:

http://cda.psych.uiuc.edu/clusteranalysis_mfiles_revised/HKS_Toolbox.pdf

If your main idea is find proper number of clusters in unlabeled images, statistical methods can be used. I can help you in this manner.

I hope it can be useful.

Regards

Hi Manav,

Unsupervised image classification is the process by which each image in a data set is identified to be a member of one of the inherent categories present in the image collection without the use of labelled training samples. Dimension reduction algorithms and clustering algorithms are the two main unsupervised machine learning algorithms needed in unsupervised image categorization. The recognition of these groups of algorithms provides a reasonable lead way into the diverse world of the application of unsupervised machine learning to image classification.

Please Refer the following Link:

https://www.cscjournals.org/manuscript/Journals/IJIP/Volume8/Issue5/IJIP-918.pdf

I think it can be useful for you.

Deep learning techniques stacked autoencoder (SAE) and Deep belief networks (DBN) can be used for unsupervised learning.