DSP Robotics Support

Yesterday I came across an interesting article on automatic texture identification using statistics. If I understood correctly, they also made it as a method for "realistic" texture creation - just from white noise and statistical cues. I like both parts. I need to focus on first one - texture identification.

So here is the article and some related stuff (I'm not matlab user, so can't tell anything about it) and sound samples.

From what I see, this should be doable in flowstone. While technical part on filters seems to be relatively clear for me, I'm completely lost with the statistics part...

So... In a few days I will try to wire it to a degree I understand it, but then I will need some help. But in the meantime - maybe someone already has more experience with it.

Okay.

Problem 1. Statistics.

// solved

Problem 2. Filter

If someone could take a look into the article it would be great. At the moment - I'm thinking about zdf bandpass filters set in paralel and maybe as butterworth or similar Q.

And for cochlear filter - I don't get that ERB part. So equation for bank of center freqs and bands according to that one would be helpful.

The modulation filterbank - output of the filter is multiplied by the center frequency that drives the filter?

*

I hope I simplified it a little bit. BTW, attaching matlab files I found in regards to that article.

(pufff...)

I think I figured out most of the generalized equations, and it's possible that... correctly. Can't tell, since I don't know what kind of values to expect, but some parameters (skewness) suggest that it may be fine.

//solved

With statistics - I was able to reach certain point, but I have no idea how to make the last one.

First part contains the hilbert transform for an array, and second - some operations on the resulting product (complex numbers on arrays part). Attaching image how it looks like.

If what I created here:
viewtopic.php?f=2&t=4363&p=24595
following this thread:
viewtopic.php?f=3&t=3175&start=10#p17132

is a hilbert transform for green arrays, then what I'm still stuck with - are the computations on complex numbers in further equations.

Statistical part is mostly done, both in ruby and greens. There are few minor issues, and I guess rounding outputs to 3 place should fix the difference problems. Simple sample arrays are here for quick checkups.

Some statistics should be normalized, but for some cases I'm not sure yet with what. But basic equations are finished.

I'm also not sure if currently used hilbert transform part - is approximation accurate enough for that kind of work.

Todo:
- filter profiles
- data handling/multiplexing scheme
- output handling (graphs)

Getting back to work.

On technical side - it's almost done.

I think interesting contribution in this project is related to statistical measurements of signals, and serialization modes for that type of processing.

Todo:

- select filter types, and configure them (freqs, Qs), and adjust destination signal paths; Martin, are you familiar with these topics mentioned in the article?

- adjust statistical parameters (cross-normalization)

- create output (either bars or replicate from article or both)

- create multistep analysis (weighted blackbox neural network?), to set identification ranges.