howdy everyone!

for the love of the raw image format, and for the love of image processing i’ve been wanting to acquire an SDK for a dSLR. at first, i was looking for the nikon SDK (since i am a nikon owner) but i’ve seen that i had to apply for it. i’ve read online that you have to be a “bona-fide developer” which i am not. i’m also not so sure what to put on my application and what i agree to as a developer if i get the sdk and agree to their terms and conditions.

so alternatively, i’ve seen online that the canon sdk is public, a GNU project. so no applications, no fuss, no rejection! which is great… but now i have to get a canon dSLR.

…oh and to my lovies, pictures are coming up soon. ^_^

I googled image segmentation

http://civs.stat.ucla.edu/Segmentation/Segment.htm

http://ieeexplore.ieee.org/iel5/4/72/00000996.pdf?arnumber=996

May need to sign on to university network to view

A few links:
http://www.dspguide.com/ch24/5.htm

http://www.owlnet.rice.edu/~elec301/Projects01/image_filt/low.html

http://www.cse.lehigh.edu/~spletzer/rip_f06/lectures/lec012_Frequency.pdf

So currently, we are trying to perform a convolution with a raw image on the TI DSP. This is really slow and not suitable for our end product (a live foreground background separation and desired background superimposition <– is that the correct word? LOL!). This is why I am looking into transforming the image (a single frame of the video) into the frequency domain and analyzing its stuff there. And that is why we have the above listed links for blalbalblalblala.

We are also having scaling problems in the TI DSP since the Y (luminosity) values are integer/char type and the matlab prototype called for type double (we had to convert the image from uint8 to double using im2double()).

EDIT

So after talking to Woo-Shik today about transforming our images to the frequency domain, it may not seem feasible to implement this method in our project. Transforming an image into the frequency domain will take a lot of computations; if you have a 100×100 image and want to transform it in the frequency domain, it would take 100 multiplications and 100 additions. He also suggested that we wouldn’t need any Gaussian smoothing and could probably just take the sobel operator and (2d) convolve it with the with the raw image. We shall see… to be continued…. DUN DUN DUNNNNN!!!

Something I wrote:

clear;
clc;

%calculate gaussian
sigma = 2.0;
size = 4;
dimension = 2*size-1;
for i = -(size-1):(size-1)
for j = -(size-1):(size-1)
k = i+size;
l = j+size;
gauss(k,l) = (1/(2*pi*sigma^2))*exp(-((i)^2+(j)^2)/(2*sigma^2));
end
end

%plot kernel
x = -(size-1):(size-1); vbbbbb
y = -(size-1):(size-1);
figure(3),surf(x,y,gauss)

img = imread(‘bw.jpg’);
img = im2double(img);
figure(1), imshow(img);

%perform convolution
for i =1:(550-dimension+1) % the number of rows in an image, the height
for j = 1:(700-dimension+1) % the number of columns in an image, the width
img2(i,j)=0;
for k = 1:dimension
for l = 1:dimension
img2(i,j) =img2(i,j)+ img(i+k-1,j+l-1)*gauss(k,l);
end
end
end
end

img2 = im2uint8(img2);

figure(2), imshow(img2);

http://amath.colorado.edu/courses/4720/2000Spr/Labs/Worksheets/Matlab_tutorial/matlabimpr.html

I’ll edit this entry and the previous entry with descriptions about the links:

http://www.cs.washington.edu/research/imagedatabase/demo/edge/

http://en.wikipedia.org/wiki/Edge_detection

http://homepages.inf.ed.ac.uk/rbf/HIPR2/log.htm

http://homepages.inf.ed.ac.uk/rbf/HIPR2/gsmooth.htm

http://homepages.inf.ed.ac.uk/rbf/HIPR2/convolve.htm

http://homepages.inf.ed.ac.uk/rbf/HIPR2/sobel.htm

http://grail.cs.washington.edu/projects/digital-matting/video-matting/

http://www.cs.brown.edu/~morgan/DefocusVideoMatting/index.htm

http://graphics.ucsd.edu/papers/camera_array_matting/

http://www.generation5.org/articles.asp?Action=List&Topic=Machine%20Vision

http://www.efg2.com/Lab/Library/ImageProcessing/Algorithms.htm