2
$\begingroup$

I have seen very educational applications of pattern recognition on complete images. I was wondering if there are methods that could handle cases with the following specifications:

  • instead of full images we have a set of sub-images which are made by tearing the original image apart
  • sub-images have arbitrary sizes (within a given range)
  • some parts of the original image might be missing after tearing it apart

I am a newbie in this field but are there (could there possibly be) convolutional networks that can handle such cases? From what I understand they work by breaking the image down.

The answer for both these cases are helpful: - Full size images for training and applying the classifier to torn-apart images - Torn-apart images for both training and testing

The answer for the both type of tasks to label each image (set of pieces) as a whole or to label each pixel in the image (~object recognition) would help me.

If you are not interested in bioinformatics don't mind this paragraph, but if you are interested, this question is the image-processing translation of my original question which is about phenotype-genotype mapping where I look at not fully assembled DNA sequences: we have contigs of reads rather than full sequences.

If you know some references that investigate this problem please advise me. Thank you very much for your input!

$\endgroup$
  • 1
    $\begingroup$ What you are looking for is called "compressed sensing". You can, for some carefully controlled classes of images (dna sounds controlled, but I can't be certain), sample (uniform random) the image for a tiny fraction of its size, and still get recognition. You have to account for pixel locations relative to each other, but it is much faster to operate on a few hundred pixels instead of tens of millions per image. link $\endgroup$ – EngrStudent Jan 25 '16 at 2:49
  • $\begingroup$ If you like, I can put it into an answer. Normally I feel like an bug among giants here, so if I am not sure what you are looking for, I feel a lot more comfortable putting the answer into a comment. $\endgroup$ – EngrStudent Jan 25 '16 at 18:06
  • 1
    $\begingroup$ Indeed you pointed me to the right direction so that counts as a right answer. I think stack exchange is about bigger bugs guiding smaller ones, thanks :D $\endgroup$ – user277194 Jan 26 '16 at 19:52
  • $\begingroup$ I think it would help if you explained the actual bioinformatical problem more carefully. From my little knowledge of it, I would think there were better analogies than image processing. Eg speech processing. $\endgroup$ – seanv507 Jan 27 '16 at 0:05
0
$\begingroup$

What you are looking for is called "compressed sensing". You can, for some carefully controlled classes of images (dna sounds controlled, but I can't be certain), sample (uniform random) the image for a tiny fraction of its size, and still get recognition. You have to account for pixel locations relative to each other, but it is much faster to operate on a few hundred pixels instead of tens of millions per image.

http://proceedings.spiedigitallibrary.org/proceeding.aspx?articleid=1354633

$\endgroup$

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.