0
$\begingroup$

I am working on a set of Images captured by an Industrial camera. However, I am not sure if I need to apply any Denoising (e.g. Gaussian or Laplacian etc) on it ?

Is there any metric that I can used to arrive at the conclusion ? Certain blogs I happened to read suggest using PSNR or MSE technique.

A sample image is attached Below. Also,my main goal is - I am passing these as input to a Deep Neural network to perform check for anomaly detection. enter image description here

$\endgroup$
1
  • $\begingroup$ please state what's supposed to be in that picture, and why it's so dark, and why you zeroed out the rest of the picture. $\endgroup$ Commented Feb 28, 2023 at 22:17

3 Answers 3

2
$\begingroup$

If it is captured by a camera I think it is fair to say that it contains noise. Shot noise as well as additive noise.

If that noise is an obstacle for your application is for you to figure out.

If denoising is a suitable preprosessor for your camera and your application is also for you to figure out. Perhaps it is better to train on noisy images directly?

$\endgroup$
0
$\begingroup$

Theoretically you don't know if you're looking at a noisy image of a perfectly even-colored object, vs. looking at a perfect image of a "speckled" object.

Realistically, you can at least to some extent use a-priori information about the object (i.e., it's smoothly shaded, it's speckled but the speckles are in a known size range, etc.). Then you can see how well the image matches that image model and declare whatever is leftover as "noise".

For anything other than quick determinations, though, that can lead to a lot of work for little gain.

$\endgroup$
0
$\begingroup$

It depends on the type of noise that you are considering. For example salt and pepper noise (mainly theoretical) is easier to detect.

  • As a metric you can consider entropy.
  • But you can also consider the color difference between adjacent pixels.
  • Another possibility is to perform a smoothing operation and compare, pixel by pixel, the differences to the original image.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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