[racket] Kernel size for flomap-gaussian-blur

From: Neil Toronto (neil.toronto at gmail.com)
Date: Tue May 14 14:02:45 EDT 2013

On 05/11/2013 01:56 PM, Dmitry.Cherkassov.dcherkassov at gmail.com wrote:
> Hi. Can't find convolution kernel dimensions in the docs, neither they can be
> specified via arguments.
> Any ideas what they are?

A kernel centered at x with standard deviation σ has its minimum at 
floor(x-3*σ) and its maximum at ceiling(x+3*σ). (Same for y and its 
standard deviation.) So the width (resp. height) is

   1 + ceiling(3*σ) - floor(-3*σ)

This can be simplified, but I took it directly from the code and didn't 
want to risk getting it wrong. It's in "images/private/flomap-blur.rkt" 
if you want to look it up.

If you want more control over convolution, you should use `array-fft' 
from `math/array'. To get good performance from that, you'll want to use 
Typed Racket, at least for writing image processing functions. Here's a 
function that converts arrays to flomaps:

   (: array->flomap ((Array Real) -> flomap))
   (define (array->flomap arr)
     (let ([arr (array->flarray arr)])
       (define ds (array-shape arr))
       (match ds
         [(vector h w)  (flomap (flarray-data arr) 1 w h)]
         [(vector h w c)  (flomap (flarray-data arr) c w h)]
         [_  (error
              "expected array with 2 or 3 dimensions; given shape ~e"

The other direction is similarly simple, because `FlArray' and `flomap' 
both store data in row-major order.

Neil ⊥

Posted on the users mailing list.