An image is taken and then rotated by an arbitrary angle and cropped. Let us assume that the original image is rather meaningless and is not aligned in any way. Given the rotated image, is it possible to learn anything about the original orientation? For example, using: 1. lossy compression artefacts (e.g. some averaging now happening in not axis-aligned rectangles), 2. any artefacts created by the camera sensors (probably camera-dependent).

  • $\begingroup$ If the image is perturbed by a faint pattern such as a grid, that pattern will survive after rotation. But won't be more visible... $\endgroup$ – Yves Daoust Sep 12 '19 at 13:47
  • $\begingroup$ @YvesDaoust If I take an arbitrary image from my phone or from the Internet, would it have any faint patterns? $\endgroup$ – Dmitry Frumkin Sep 12 '19 at 14:02
  • $\begingroup$ If what you mean is loss of data due to rotation, to some extent you can guess original rotation, but, 90, 180, 270 degrees changes nothing and any rotation will get mirror cases that are not distinguishable. The artifacts (in case of say rectangle) are technically not compression, but loss of data, which cannot be represented, so it is effectively lossy compression in some sense. Artifacts from camera came from the same problem - putting seemingly continous image into discrete grid. It is hard to tell apart blurred rectangle (prior to rotation) from rotation artifacts. What is your goal? $\endgroup$ – Evil Sep 12 '19 at 14:07
  • $\begingroup$ @Evil The goal is given a rotated image to learn the orientation at which it was taken using either lossy compression or camera-dependent artefacts. For example, I see an image that somebody aligned and want to learn its orientation before the manual alignment. Or at least the approximate orientation, or, at a minimum, whether rotation was performed. $\endgroup$ – Dmitry Frumkin Sep 12 '19 at 14:14
  • $\begingroup$ @DmitryFrumkin: strongly compressed JPEGs can be recognized, you see the grid. JPEG2000 and losslessly compressed images can not. $\endgroup$ – Yves Daoust Sep 12 '19 at 18:13

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.