2

I'm implementing a Gaussian filter in which, as explained in one Dimensional gauss convolution function in Matlab, I'm normalising the filter coefficients by dividing them by sum of all the coefficients. Normalizing by max value or by total value? explains that normalisation can be done by dividing by maximum of the coefficients too.

I'm unable to get the point that what should be the conditions/specifications to be considered in choosing the normalising factor?

Sumbul
  • 167
  • 1
  • 8

1 Answers1

1

Conventionally, you should divide by the sum of all the coefficients, so that the DC gain is unity (i.e. low frequency signals are not changed by the filter).

NickJH
  • 561
  • 3
  • 7
  • 1
    and when should it be divided by maximum value? – Sumbul Nov 15 '17 at 12:41
  • I can't think of a reason to divide by the maximum. [There exist some fixed-point systems that can most easily mutiply by numbers in the range [-1..1) and on those you might scale the largest coefficient magnitude to unity, for precision. But that doesn't apply in Matlab, which uses floating-point.] – NickJH Nov 15 '17 at 13:59