0
$\begingroup$

Is there any way to bring different classes of spectrograms to comparable amplitude levels so that when they are used for classification, the deep learning algorithm focuses on other aspects (like the presence or absence of harmonics)?

Thanks in advance.

$\endgroup$

1 Answer 1

0
$\begingroup$

In most spectral classification applications, you want to whiten the spectrum such that the noise is flat and at a consistent amplitude between spectra. Ideally, you would have some knowledge about what transform your signal is going through and you would apply an inverse of that transform to account for frequency-dependent attenuation/gain. If you don't know that transform, there is still hope! Using the assumptions that your signals of interest are narrow band and your noise is broadband, you could normalized by an an order statistic filter applied along frequency with something like a 10th percentile and a filter length several times the maximum bandwidth of your signal of interest. That will essentially normalize your spectrum to the noise, giving your classifier a reasonably flat, consistent background to extract out signal features.

$\endgroup$

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.