Thread Subject: Find Overlap of Two Images

Hi,

My question applies for image stitching.

You can look at two images I have in mind here, bigsur1.jpg and bigsur2.jpg

http://www.cb.uu.se/~cris/blog/index.php/archives/76

What I would like to do is to be able to find the region between the two images that overlap.

Specifically I want to find the sum of squared difference between the overlap in first image and overlap in the second image.

Thanks for your help.

"omegayen " <omegayen@ameritech.net> wrote in message <hcgatu$44m$1@fred.mathworks.com>...
> Hi,
>
> My question applies for image stitching.
>
> You can look at two images I have in mind here, bigsur1.jpg and bigsur2.jpg
>
> http://www.cb.uu.se/~cris/blog/index.php/archives/76
>
> What I would like to do is to be able to find the region between the two images that overlap.
>
> Specifically I want to find the sum of squared difference between the overlap in first image and overlap in the second image.
>
> Thanks for your help.

There is a similar thread here: http://www.mathworks.com/matlabcentral/newsreader/view_thread/259867

I run the code on the two images of this webpage. For convenience I copy the code bellow.

% Bruno

function PatternMatching(GPU)

%% Test data
n = 0; % 0, 100, 500, 1000
switch n
    case 0,
        im1=imread('bigsur1.jpg');
        im2=imread('bigsur2.jpg');
        A = sum(double([im1; im2]),3);
        A1 = A(1:size(im1,1),:);
        A2 = A(size(im1,1)+1:end,:);
    otherwise
        error('Please setup crop-size for n=%d', n)
end

if n==0
    maxshift = [100 700];
end

% Use GPU flag
if nargin<1
    GPU = true;
end

%%
tic
% should return the same input as above
[dx dy] = pmatch(A1, A2, maxshift, GPU);
toc
fprintf('Shift found [pixel] is (%d,%d)\n', dy, dx);


%%
% Graphic check

fig=figure(1);
clf(fig);
ax = axes('Parent',fig);
x1 = 1:size(A1,2);
y1 = 1:size(A1,1);
hold(ax,'on');
if n==0
    A1=im1;
    A2=im2;
end
imagesc(x1,y1,A1,'Parent',ax);
plot3(ax,x1([1 end end 1 1])+[-1 1 1 -1 -1]/2,...
         y1([1 1 end end 1])+[-1 -1 1 1 -1]/2,...
         ones(1,5),'k');
x2 = dx + (1:size(A2,2));
y2 = dy + 1:size(A2,1);
imagesc(x2,y2,A2,'Parent',ax);
plot3(ax,x2([1 end end 1 1])+[-1 1 1 -1 -1]/2,...
         y2([1 1 end end 1])+[-1 -1 1 1 -1]/2,...
          ones(1,5),'k');
% Intersection
x = intersect(x1,x2);
y = intersect(y1,y2);
plot3(ax,x([1 end end 1 1])+[-1 1 1 -1 -1]/2,...
         y([1 1 end end 1])+[-1 -1 1 1 -1]/2,...
          ones(1,5),'r','LineWidth',1);
set(ax,'Ydir','reverse');
axis(ax,'equal')
 
end % PatternMatching

%%
function [dx dy] = pmatch(A1, A2, maxshift, GPU)
% function [dx dy] = pmatch(A1, A2, maxshift, GPU)
% Pattern matching engine

% Use GPU
if nargin<3 || isempty(maxshift)
    % We don't look for shift larger than this (see ImageAnalyst's post)
    % half of the size of A1 and A2
    maxshift = ceil(0.8*min(size(A1),size(A2)));
end

% Use GPU
if nargin<4 || isempty(GPU)
    GPU = true;
end

if isscalar(maxshift)
    % common margin duplicated for both dimensions
    maxshift = maxshift([1 1]);
end

% Select 2D convolution engine
if ~isempty(which('convnfft'))
    % http://www.mathworks.com/matlabcentral/fileexchange/24504
    convfun = @convnfft;
    convarg = {[], GPU};
    fprintf('GPU/jacket flag = %i\n', GPU);
else
    % This one will last almost forever
    convfun = @conv2;
    convarg = {};
    fprintf('Slow Matlab CONV2...\n');
end

% Correlation engine
A2f = A2(end:-1:1,end:-1:1);
C = convfun(A1, A2f, 'full', convarg{:});
V1 = convfun(A1.^2, ones(size(A2f)), 'full', convarg{:});
V2 = convfun(ones(size(A1)), A2f.^2, 'full', convarg{:});
C2 = C.^2 ./ (V1.*V2);
center = size(A2f);
C2 = C2(center(1)+(-maxshift(1):maxshift(1)), ...
        center(2)+(-maxshift(2):maxshift(2)));
[cmax ilin] = max(C2(:));
fprintf('Correlation = %g\n', sqrt(cmax))
[iy ix] = ind2sub(size(C2),ilin);
dx = ix - (maxshift(2)+1);
dy = iy - (maxshift(1)+1);

end % pmatch