Archive: ConvoKernels


25th December 2005 01:06 UTC

ConvoKernels
Here are some convolution kernels for y'all to use at will in your presets. They were originally intended for still pictures, but work just fine in the trans/convolution for AVS. Some have odd/interesting/unpredictable outcomes. I would hit the "auto" button before enabling it just to get a baseline for any given one, and the 2-pass, absolute and wrap mostly all have interesting variations as well. Many are standard DSP kernels and all are labelled with the function that comes of using it with a static bitmap. So have fun playing with them, 69 in all.


25th December 2005 02:05 UTC

Very, very nice.

Thanks!


25th December 2005 05:15 UTC

ya cool dude! Thanks!


26th December 2005 23:40 UTC

I need to emend this post, for since I've found that these files do not exactly correspond to the function the file name seems to indicate. They still work as convolution kernels, but just not as the file name suggests.

All this came out of some studoes I've been undertaking regarding DSP, and the kernels themselves came from Michael Vinther's Image Analyser. They originally had a .FIR extension, and it wasn't until yesterday that I noticed that the numbers originally entered within IA fail to come up the same way when loaded into convolution.ape. Why this is, I do not know, sonce the format is the same; probably having to do with a byte offset or something. <sigh>

So the bottom line is that though they are usable, the outcomes are unpredictable (which still can be fun). I am working on making true function standardised kernels myself based upon matrices that I copied long ago, and these will be hand-entered and saved in native convolution.ape format, and from these one will expect a more predictable outcome. Might take me a month or so, given that I have 10-12 pages of matter to go through, and how slow it is to actually do a bunch of em manually. I will post them when ready to this thread though.


27th December 2005 00:03 UTC

Since the thread exists....

These are some common convolution filters that I created a while ago in matlab, they are all mathematically correct (other than quantisation since the avs convo is integer)


27th December 2005 11:31 UTC

how can you fuck up with convolution filters... they are so painfully basic.

/me slaps hornet

dont post crap :P


27th December 2005 15:06 UTC

Call me stupid, but how exactly does the convofilter work? Does it take every pixel and add to the are around it the amount you tell it to, or what?


28th December 2005 03:04 UTC

bingo. pretty much what you said, but for each pixel it takes its own colour multiplied by the number you put in the relevant convo filter text box and divides it by the divisor (scaling). bias adds a constant to everything before the dividing happens... it then adds them all up in another buffer to produce the 'convolved' image

there is a nasty 'mathematically perfect' analogue of this which is actually called 'convolution', but i recommend you dont ever look it up, it wont help you in avs.

its kinda like how the simple factorial (n!=n*(n-1)*(n-2)...*3*2*1) has a more flexible analogue in the gamma function (involves exponentials and integration), but you dont need to know about gamma to use factorial.


29th December 2005 15:29 UTC

Now this raises another question. Is the color of a pixel one number (0..2^BitRate) or a bunch of numbers, blue, green, red, hue, saturation, that other one, brightness, etc.?


30th December 2005 02:18 UTC

The Scientist and Engineer's Guide to Digital Signal Processing is very helpful in understanding all this stuff, particularly chapters 23, 24 and 25. Its available at http://www.DSPguide.com/ and is free.


30th December 2005 15:58 UTC

Wow. Very nice. An entire book for free. Wow.
A little old, it makes a mention of TV being impossible to transmit via a dialup connection, but whatever. Thanks.