Cross Stitch to Matlab

Question

My girlfriend was saying that she may start doing some Cross Stitching and that she may want to make a drawing, image or whatever. In that moment I though "Well, I should be able to create a Matlab code to get any image and convert it to a cross-stitch patter". It turns out I am not the first one to think about this.

But Im not sure if I am doing the right thing.

So lets illustrate an example:

Suppose a pixelated image, any size, any color palette. for example the following screenshot from pixel artist WANEELLA

Suppose that we don't want to scale the image, the result image will have the same amount of pixels as the original (else a imresize will do).

Now the problem is using only the color palette from the available ones. I decided to use the DCM palette, mainly because I found the RGB conversion of it.

I created away of colour quantization. I use Lab colors to find the closest colour in the DCM palette and use that one.

clear;clc;
% read image (its a gif)
[img,C]=imread('wn.gif');
% Convert it to RGBimage.
img2=img(:,:,:,3);

imgC=zeros([size(img2) 3]);

for ii=1:size(img2,1)
    for jj=1:size(img2,2)
        imgC(ii,jj,:)=C(img2(ii,jj)+1,:);
    end
end
img=imgC;
imshow(img)

% read DCMtoRB conversion
fid=fopen('DCMRGB.txt');
fgets(fid);
ii=0;
tline = fgets(fid);
while ischar(tline)
    ii=ii+1;
    table{ii}=tline;
    tline = fgets(fid);
end
fclose(fid);

for ii=1:size(table,2)
   DCMRGB(ii,1)=str2num(table{ii}(1:4));
   DCMRGB(ii,4)=hex2dec(table{ii}(end-5:end-4));
   DCMRGB(ii,3)=hex2dec(table{ii}(end-7:end-6));
   DCMRGB(ii,2)=hex2dec(table{ii}(end-9:end-8));
end

% origColous=reshape(img, [], 3);
Colours=double(unique(reshape(img, [], 3), 'rows'));
Ncol=size(Colours,1);



cform = makecform('srgb2lab');
DCMLab = applycform(DCMRGB(:,2:4)./255,cform);
Colourslab = applycform(Colours,cform);


eudist=@(p)sqrt(p(:,1).^2+p(:,2).^2+p(:,3).^2);
Cind=zeros(Ncol,1);

for ii=1:Ncol
    aux=ones(size(DCMLab,1),3);
    aux(:,1)=Colourslab(ii,1);
    aux(:,2)=Colourslab(ii,2);
    aux(:,3)=Colourslab(ii,3);
    d=eudist(DCMLab-aux);
    [~,Cind(ii)]=min(d);
end
% now DCMRGB will have  DCMcode, R, G, B
% Perform map conversion
img2=zeros(size(img));
indimg=zeros(size(img,1),size(img,2));
for ii=1:size(img,1)
    for jj=1:size(img,2)
        %wich colour is the pixel?
        [~,indx]=ismember(double(squeeze(img(ii,jj,:)))',Colours,'rows'); 
        indimg(ii,jj)=Cind(indx);
        img2(ii,jj,:)=DCMRGB(Cind(indx),2:4);
    end
end


%%
subplot(121)
imshow((img))
% subplot(222)
% [X_dither,map]=rgb2ind(img,DCMRGB(:,2:4)./255,'nodither');
% imshow(uint16(X_dither),map);

subplot(122)
imshow(double(img2)./255)

The result looks like:

However, in this webpage : http://www.picturecraftwork.com/ As you can see, the choice of colours is different in the webpage, and actually it makes more sense and it gives a quite good feeling even without the real colormap.

After a lot of going around, I believe that there could be 2 main things I need to change and implement.

1.- I dont trust 100% the DCMRGB values. I emailed the company seeking more information about theyr color palette.

2.- The brigthness, contrast, hue and saturation values have enormous influence in the output.

how can I modify this 4 values (as in the webpage) in the desired image using Matlab?

---

---

DMC2RGB file: (I can copy paste it here if needed)

http://pastebin.com/qixUgnvy

DCM color pallete:

Answer 1

Okay, here's a generalised suggestion, although I haven't managed to get great results with it yet.

1) Start by reducing your original image to a sensible number of colours using rgb2ind with nodither, specifying only the number of colors (not the map used).

[I_ind, old_map] = rgb2ind(I,64,'nodither);

You can experiment with the number of colours for a given image that fit without losing too much detail. Basically you want to figure out how many unique colours you want first (seeing as you're cross-stitching the output, you don't want a bunch of colors that only are used for about three stitches each).

2) Convert both maps (the one that's the output of rgb2ind and the one that you load from the dcm to rgb file) to an appropriate colour space (I used HSV because I was feeling lazy, L*a*b is probably better).

3) Select colours by checking which indexed colours have the most pixels present, picking the one that's closest, then removing that colour from the candidate list so we end up with the 64 or whatever

% use histogram and sorting - assign colours to most common values first
[plist pbins] = hist(I_ind(:),0:63);
[plist_sorted, sort_ind] = sort(plist, 'descend');

% old_map = the output of rbg2ind
% new_map = the new one we're about to make
% dcm_map = (copy of!) the dcm map in the same colorspace 
new_map = zeros(size(old_map)); 

% loop through from most to least common, remove a colour from the dcm map if we've used it.
for n = 1:sort_ind
   D = pdist2(old_map(n,:),dcm_map);
   m = find(D==min(D),1,'first');
   new_map(n,:) = dcm_map(m,:);
   dcm_map(m,:) = [];
end

I used the image you used to have up on the question. As it comes out of rgb2ind:

As it comes out at the end with 64 DCM colors (a bit off, probably because my choice of comparison is wrong, but guaranteed to have 64 separate dcm colors).

I guess what you need to find is the right weighting of the different factors (you can put a weighting function into pdist2 quite easily). For example, weighting HSV as [0.8, 0.1, 0.1] gave me this brilliant madness: