Reply by Fred Marshall December 11, 20082008-12-11
clay@claysturner.com wrote:
> On Dec 10, 1:53 pm, Rune Allnor <all...@tele.ntnu.no> wrote: >> On 10 Des, 19:41, vijay <jahagirdar...@gmail.com> wrote: >> >>> Newbie Here. >> >>> I have been told that and we normally go for 3,5,7,9 Tap poly phase >>> filters for image processing.... and 4,6,8 Tap filters should not be >>> used... >>
Your "Subject" line and the body of your question refer to two different things. Everything that Rune and Clay have mentioned are good points and Clay touched on the Subject line briefly. That is, the Subject mentions polyphase filters. Generally, each "phase" of a polyphase filter has a number of regularly-spaced zero-valued input samples (which need not be multiplied by any coefficient - thus leading to the polyphase notion). Might that lend some insight? Fred
Reply by December 10, 20082008-12-10
On Dec 10, 1:53&#2013266080;pm, Rune Allnor <all...@tele.ntnu.no> wrote:
> On 10 Des, 19:41, vijay <jahagirdar...@gmail.com> wrote: > > > Newbie Here. > > > I have been told that &#2013266080;and we normally go for 3,5,7,9 Tap poly phase > > filters for image processing.... and &#2013266080;4,6,8 Tap filters should not be > > used... > > That's because with image processing, one usually wants to > associate the computed quantity with an existing pixel. > With a 3x3 mask (view with fixed-width font), > > M = [ a b c > &#2013266080; &#2013266080; &#2013266080; d e f > &#2013266080; &#2013266080; &#2013266080; g h i] > > one centers the mask on a pizel (n,m) and store the computed > value in the corresponding position in the new image. Try the > same with a 2x2 mask: > > m = [a b > &#2013266080; &#2013266080; &#2013266080;c d] > > Which one of these coefficients do you center on (n,m)? > You can choose arbitrarily, but the computed result will > depend on your choise. So there is a lot of hazzle that > just goes away when you use odd x odd dimensions for the > masks. > > And of course, lots of the common operators are symmetric > around the center, like the Gaussian mask and various > gradient operators. > > Rune
To expand on what Rune stated, linear phase filters must have either a symmetric or antisymmetric impulse response. Also all linear phase FIR filters have a delay of (N-1)/2 taps where "N" is the length of the impulse response. You might not want to shift your image over by half of a pixel. So odd lengths allow you to easily compensate for the overall integral number of samples delay imposed by an odd length filter. Not so easy with an even length filter. Clay
Reply by Rune Allnor December 10, 20082008-12-10
On 10 Des, 19:41, vijay <jahagirdar...@gmail.com> wrote:
> Newbie Here. > > I have been told that &#2013266080;and we normally go for 3,5,7,9 Tap poly phase > filters for image processing.... and &#2013266080;4,6,8 Tap filters should not be > used...
That's because with image processing, one usually wants to associate the computed quantity with an existing pixel. With a 3x3 mask (view with fixed-width font), M = [ a b c d e f g h i] one centers the mask on a pizel (n,m) and store the computed value in the corresponding position in the new image. Try the same with a 2x2 mask: m = [a b c d] Which one of these coefficients do you center on (n,m)? You can choose arbitrarily, but the computed result will depend on your choise. So there is a lot of hazzle that just goes away when you use odd x odd dimensions for the masks. And of course, lots of the common operators are symmetric around the center, like the Gaussian mask and various gradient operators. Rune
Reply by vijay December 10, 20082008-12-10
Newbie Here.

I have been told that  and we normally go for 3,5,7,9 Tap poly phase
filters for image processing.... and  4,6,8 Tap filters should not be
used... I am not able to find any reference online to understand the
issues with even tap polyphase filters. Can someone point me to a
suitable online reference.

Regards
Vijay