0
$\begingroup$

Abstract Problem: I have several series (sharing the same class) of about 400 small images each. Each one of my series has >390 images, that belong together. The rest of the images shares the style of its series, however may include something, which makes it look odd from the human eye perspective.

Is it possible to detect, if a new series of images contains an odd looking image?

Example:

You have people and take 400 high resolution images of their skin of their arm. The image do not overlap. Most images (>390) will not show anything interesting, however some may indicate a desease, that will be noticeable, especially if it's repeating across multiple people. I don't know, which of these images looks odd, but I know, if a person has a desease or does not have one.

Thus far:

My first idea was to use some sort of cluster algorithm. It is unsupervised and thus functional for the problem. Via transfer learning I create a good feature space of my images, maybe reduce the dimensionality of it, cluster them and hope there will be 2 well seperated kernels, that contain 10 or about 390 samples each. This method also has another advantage: It can be done fairly quickly, no training of some model etc. is necessary.

My second idea would be to cross compare all images inside a series, but overall it would follow a similar path from my 1st idea. The hope is that the difference between two not-odd images is fairly boring, while comparing odd and not odd is noticeable.

3th: Use some outlier detection using Autoencoders/GANs.

Question: Is there a deep learning approach for this type of problem? At the moment all my ideas revolve around one single series, however all series belong to the same class - which has to be somehow useable. Is someone familiar with this kind of problem? Are Odd-one-out networks is useable/applicable in this situation? To my knowledge however, they are self-supervised and thus maybe not the right tool.

I would be really happy, if someone with more experience than me, could point in one reasonable direction and recommend me something for further reading. Thanks in advance.

$\endgroup$
1
$\begingroup$

This is the classic case of "local outlier detection". You have a set of images, but a few objects divert a little bit more than the others.

There exist some deep learning approaches for this, but they seem to mostly work for images (read: they work on MNIST). As I don't use images, I have not further studies them.

As a starting point for local outlier detection prior to deep learning, see this overview that discusses some applications (there is video in there, but IIRC not deep learning):

Schubert, E., Zimek, A., & Kriegel, H. P. (2014). Local outlier detection reconsidered: a generalized view on locality with applications to spatial, video, and network outlier detection. Data Mining and Knowledge Discovery, 28(1), 190-237.

$\endgroup$
  • $\begingroup$ This is already a huge help!! Thank you very much. Once a direction is given, everything else will follow from that. $\endgroup$ – Imago Nov 25 '19 at 14:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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