yuli nurachman

Subject:	Re: skin detection	Actions...
From:	ber202_85 (ber2...@hotmail.com)
Date:	Oct 11, 2008 5:17:33 am
List:	com.googlegroups.aforge

Hi Andrew

Thanks for your reply. I like your suggestion although I'm not convinced. The thing is that the method I used - the YCbCr had thresholds which were based on some ranges like you're saying.

http://w3.ualg.pt/~ftomaz/download/dicta2003.pdf

Thats the paper I read and which I think is the most accurate. Some of the equations I can't understand - the Ms, Mt and sigma - I dont know how to get their values.

If you think though that your method is better, I try and implement it - however if you think that the method in that paper I posted is better I'd rather try and understand it and implement it cos as I said before I need accurate results.

Thanks for your help.

Best Regards Bertrand Azzopardi

On Oct 11, 7:34 am, Andrew Kirillov <andr...@gmail.com> wrote:

Hello,

I found this paper

Which one? Do you have a link?

Most of the papers I found use matlab and I don't know matlab.

This is quite common. You don't need to know Matlab a lot. All you need is to understand idea from the paper and get equations.

You may try this … Take some (for example 50) pictures of people/ faces. Edit them in any editor and remove all non-skin regions, so pictures will have only skin regions. Then use AForge.NET, for example, and try find statistics about skin ranges in different color spaces (RGB, HSL, YCbCr, for example). When you get skin ranges for different color spaces, try apply them to real (source) images and see what is the amount of false positives. Take the method, which perform best for you, if any. This would be a good start, which is based on some small investigation and statistics. And this is method used in some papers …

With best regards, Andrew

On Oct 10, 9:29 pm, ber202_85 <ber2...@hotmail.com> wrote:

Hi Andrew

I found this paper and i just implemented it. But im not happy with the results. I'm using Hough transforms to detect the face so the preprocessing stage of skin detection is very very important. I need some algorithm which is accurate meaning it would eliminate as much as false positives as possible.

I cant understand how to find the covariance and the mean values using the Em algorithm. If you could kindly provide me with some explanation or some source code it would be more than grateful. The thing is that I'm implementing the face detection system using c# so all the code has to be written from scratch. Most of the papers I found use matlab and I don't know matlab. Please let me know if you find something suitable for my application.

Best Regards Bertrand Azzopardi

On Oct 10, 5:47 pm, Andrew Kirillov <andr...@gmail.com> wrote:

Hello,

I would like to ask you some questions regarding your current implementation: 1) Why did you choose YCbCr color space for locating skin regions? 2) How did you determine that your current ranges for each component (Y, Cb, Cr) are optimal?

As you know, there are many different papers on the topic of skin detection, so I am just wondering what is the idea behind your approach? Is it based on some known research (or unknown - maybe you’ve done something) or did you just selected these ranges because of … what?

With best regards, Andrew

On Oct 10, 4:12 pm, ber202_85 <ber2...@hotmail.com> wrote:

Hi Andrew

I'm building a face detection system in c# for my dissertation using hough transforms. I'm getting false positives in my results and this is due to the fact the skin detection process is not accurate. The method I'm using at the moment is the following:

private void YCbCr(Bitmap bmp) { double Y = 0.0; double Cb = 0.0; double Cr = 0.0;

for (int x = 0; x < y =" 0;" c =" bmp.GetPixel(x," y =" (0.257" cb =" (0.148" cr =" (0.439">

if ((c.R > 95 && c.G > 40 && c.B > 20 && (Math.Max(Math.Max(c.R, c.G), c.B) - Math.Min(Math.Min(c.R, c.G), c.B) > 15) && (c.R - c.G) > 15 && c.R > c.G && c.R > c.B) && (Cr > 140 && Cr <> 105 && Cb <>

else { bmp.SetPixel(x, y, Color.FromArgb(0, 0, 0));

} } } }

As you can see it is not so efficient. Do you know of any other method maybe provide me with some source code cos I'm stuck.

I was reading about the Mahalanobis distance but I can't figure out how to retrieve the mean values and the covariance. If you could help me out it would be greatly appreciated.- Hide quoted text -

- Show quoted text -- Hide quoted text -

- Show quoted text -

yuli nurachman

Kamis, 01 Oktober 2009

Tidak ada komentar:

Pengikut