I am interested in studying local pixel value dependencies in images and I am wondering if it exists a general form of prediction-error convolutional filter that is able to suppress the image's content and extract the residual between the central pixel value and it's local neighbors within the filter dimension.

If so, what kind of filters I can use and do they have a general form or structure in their weights.

  • $\begingroup$ you mean, like, the differences between a pixel and it's neighbors? $\endgroup$ Oct 27 '16 at 21:24
  • $\begingroup$ Yes, I'm interested in filters that generate white noise for instance to visualize discontinuities in images. It's sort of predicting linear combination of the central pixel in terms of its neighbors locally. $\endgroup$
    – user2987
    Oct 27 '16 at 21:28
  • 1
    $\begingroup$ you're describing your filter kernel in your question; it's a simple $$\begin{pmatrix}0&-1&0\\-1&4&-1\\0&-1&0\end{pmatrix}$$, for example, and all these filters are typically rather boring high-pass filters (as you notice yourself: they detect discontinuities). $\endgroup$ Oct 27 '16 at 21:30
  • $\begingroup$ I tried already high pass filters which preserve high variation in images and that gave me sort of edge maps. I am interested to have an output that looks like a noise field where the content is completely suppressed. Does it exist such convolutional filter? $\endgroup$
    – user2987
    Oct 27 '16 at 21:38
  • $\begingroup$ @MarcusMüller I just tried the simple example that you suggested and it looks that it generates what I need. Do these filters have certain general form so that I can use many kinds of them and compare. Like having -1 around the central pixel and filter coefficients sum up to 0. $\endgroup$
    – user2987
    Oct 27 '16 at 21:46

Your Answer

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

Browse other questions tagged or ask your own question.