2

Some of my algorithms rely on segmenting an image based on the count of significant horizontal lines detected, images taken of computer screens or monitors can be problematic because the distortion present results in poor binarization and unclearly defined areas.

Camera capture of monitor: capture of monitor

The ideal situation is for the camera captured image to be rendered without the distortions resulting from capturing through the monitor (as it appears digitally), but is it possible to detect that an image is captured in this manner using opencv, and therefore reject it if that is the case(as it cannot be processed)?

Avan
  • 223
  • 3
  • 13
  • 2
    The phrases "some of my algorithms" and "an image" are quite vague. Are they "OCR algorithms" for "music and text images" ? Are they "face recognition algorithms" for "selfies" ? Help us to help you -- spell out what you hope to accomplish and what road blocks are currently preventing that. – J_H Apr 03 '21 at 02:08
  • 2
    This is not trivial. How would you define such noise? One very basic (and incomplete, surely) way would be that the noise seems to be periodic. It has some amplitude and frequency (or frequencies). Maybe try inspecting the image in the frequency domain and look out for suspicious bursts of gain in the spectrum. Maybe check out the Power Spectral Density while you are at it and see if the noise can be identified there as well. – stateMachine Apr 03 '21 at 02:15

1 Answers1

5

A sensor (camera) discretizes an observed continuous scene into distinct pixels. The troublesome image sources you cite, LCD / OLED digital displays and CRTs, discretize a scene into their own pixels, so we have double discretization, which beat against one another. This accounts for the annoying moiré patterns visible in your example image.

Apparently you wish to detect such patterns. Consider using FFT. Pick a segment length L, and a segment center location. Then rotate a line through several angles, reading pixels from the segment and handing them to FFT. A moiré pattern will exhibit periodicity for at least one angle. And nearby segment centers, at same angle, will exhibit similar periodicity.

Perhaps this is a solved problem? https://github.com/AmadeusITGroup/Moire-Pattern-Detection https://github.com/AmadeusITGroup/Moire-Pattern-Detection/blob/master/src/positiveImages/355_letterbox1024.jpg

But perhaps "rejecting" isn't quite your true goal at all. Tell us about the failure mode, about what goes wrong when your algorithms encounter images from troublesome sources. Could we maybe deal with the discretization by running a light gaussian blur over the image? Possibly followed by unsharp masking?

You didn't mention whether you have any control over how the images are acquired. Obtaining a pair of images of "same scene", separated in time by a second or so, would offer you lots more information. If sensor is tripod mounted, consider moving it a millimeter or so before snapping the 2nd image.

J_H
  • 17,926
  • 4
  • 24
  • 44