I have a set of grayscale images, some of them are transformed of the other images. For example in 10 images, image 2 is the same as image 8 but rotated, and image 4 is the same as image 7 but translated. There might be a slight distortion but they look mostly the same. And lastly, there could be more than two copies of the same image, but it would be the exact same shape with slight distortion (nothing like cats and dogs).

I tried the image similarity using historgram but it can lead to very large erros and is not reliable. Also convolutional neural networks need to be trained beforehand. I'm looking for an unsupervised method because the data are not labeled, different datasets have different contents and the number of images are too small for a neural network.

I don't expect you to explain the whole method (I would appreciate it though), just looking to find a direction to do my own reading.

  • 1
    $\begingroup$ There is plenty of literature on image similarity search, using many different techniques, not only histograms. There are also keypoint based approaches, for example, that try to be rotation invariant. Just get a book. $\endgroup$ – Has QUIT--Anony-Mousse Aug 18 '18 at 7:31
  • $\begingroup$ @Anony-Mousse the summation of your comment: there are many methods, get a book. $\endgroup$ – anishtain4 Aug 19 '18 at 0:06
  • 2
    $\begingroup$ Sorry. I cannot copy & paste an entire book into a comment... Your question is too broad. $\endgroup$ – Has QUIT--Anony-Mousse Aug 19 '18 at 8:57
  • 1
    $\begingroup$ Sorry, I don't do image analysis, so I don't have an exact book title to recommend. But I've looked up things in these books, so I know that there are keypoint based approaches that were the state of the art before deep learning for most image analysis (including unsupervised). $\endgroup$ – Has QUIT--Anony-Mousse Aug 20 '18 at 15:50
  • $\begingroup$ @Anony-Mousse I looked up keypoint based approaches and they'll do what I'm looking for. You can post it as an answer so I accept it. $\endgroup$ – anishtain4 Aug 21 '18 at 14:40

SIFT key points are a starting point, but there are many more.

Scale invariant feature transform

Lowe, David G. (1999). "Object recognition from local scale-invariant features" (PDF). Proceedings of the International Conference on Computer Vision. 2. pp. 1150–1157. doi:10.1109/ICCV.1999.790410.


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