how to find bitwidt for representing gaussian in hardware

Seprability of gaussian should be implemented, but how to calculate the perfect 1d kernels for gaussian 33x33 size and sigma were k=0-5

Answer 1

[Aquib Rashid]

We find the convolution of double precision image(created by rand function) with double precision gaussian.....(A) Then find the convolution of quantized image with quant gaussian....(B)

Subtracting A and B for different bitwidth will give the results, as after around 15 word length, the error remains constant, this is the quantized error which remain, coz of the quantization in the image.