1
$\begingroup$

When building deep learning models for image analytics-related applications, we sometimes apply various types of operations to enhance the image, such as an image denoising operation.

In my study, we have images generated by physical simulations. In other words, the physical simulations generate matrices of dimensions such as 256*256, which can be visualized as an image as well.

I am trying to apply a deep learning model to perform some analysis over these physics-based images. In the pre-processing steps, I can always apply those image analytics related techniques to pre-process my images, but I am not sure whether it makes sense. For example, denoising or some other contrast enhancement operations can be used to improve the quality of images, like photos. But would it make sense to use them process the images generated by physics simulation?

$\endgroup$
  • 3
    $\begingroup$ Can you post an example of an image generated by these physical simulations? $\endgroup$ – timleathart Jan 7 '17 at 0:13
  • $\begingroup$ If you're expecting/experiencing noise in your images which were generated by a simulation, then that is most probably caused by the technique you're using to simulate your physical model. In that case, instead of worrying about denoising the images, it is a better option to work with the simulation, so that they don't produce any noise. If that's not possible for you, then you should at least know the reasons of the noises, and should expect the points at which the noise can be produced. Once you're done with estimating the location of the potential noises, a median filter should work great. $\endgroup$ – Syed Ali Hamza Jan 7 '17 at 19:58
  • $\begingroup$ The most common use of an ML model such as CNN is to convert images into a model based on statistical regularities in the observations. You already have a model (the physical simulation), so could you clarify what you are trying to achieve? A common matching scenario would be that you hope to train an NN on simulation (where you know params) in order to later train and predict params of a theoretical model from real-world images. Is that the case here? Understanding whether you are doing this or something different is key to advising you how to process the images. $\endgroup$ – Neil Slater Jan 9 '17 at 10:09
0
$\begingroup$

I think your intuition is correct, that it doesn't make sense to "clean" the matrices, if they are generated perfectly. I agree with @SyedAliHamza that noise should be eliminated as much as possible in the simulation.

On the other hand, depending on what analysis you are trying to do, you may want your model to generalize better. In that case, it would be good to apply some data augmentation, meaning adding some noise or other types of augmentation to your matrices.

|improve this answer|||||
$\endgroup$

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.