Techniques for blending images that have overlap

48 visualizzazioni (ultimi 30 giorni)
I have some images that all overlap each other by a known amount (e.g. 15% of the pixels on along a side). No registration should be needed since I know exactly the overlap amout, i.e. placement. My method of merging them into a single image is by taking the ratio of the mean values of the images' overlapping regions and then correcting one of the images by that ratio. For example:
ratioAB = mean(matA_ovlap(:))/mean(matB_ovlap(:));
matB_new = matB*ratioAB;
This is better than simply piecing them together, but still clearly leaves a lot to be desired. See result below. Any recommendations for a more sophisticated blending?
  2 Commenti
Paul Safier
Paul Safier il 15 Set 2021
@KSSV what would the weights be? The distance within the overlap region?

Accedi per commentare.

Risposta accettata

DGM il 14 Set 2021
Modificato: DGM il 14 Set 2021
This is a simple approach to doing edge-interpolation. This could also be done using interpolation tools in fewer lines, but right now I don't have the brain cells to spare in any effort to make this elegant. Just bear in mind that there's plenty of room for improvement. Anyone is free to post something better.
For this example, I'm going to use RGB color images. That might not be what you're using.
tiling = [3 3]; % image tiling [nup nacross]
overlap = [64 64]; % in pixels [rows cols]
% generate colored test images as a 4D RGB frame stack
% some of this uses MIMT; ignore that.
nframes = prod(tiling);
inpict = imread('cameraman.tif'); % raw single-channel source
A = zeros([imsize(inpict,2) 3 nframes]);
A(:,:,:,1) = repmat(im2double(inpict),[1 1 3]).*permute([1 0.3 0.8],[1 3 2]);
for f = 2:nframes
A(:,:,:,f) = imtweak(A(:,:,:,1),'lchab',[1 1 (f-1)/nframes]); % rotate hue
% do setup
overlap = round(overlap/2)*2; % make sure it's even
s0 = [size(inpict,1) size(inpict,2)]; % source frame geometry
st = 2*(s0-overlap) + (s0-2*overlap).*(tiling-2) + overlap.*(tiling-1); % output image geometry
myn = repmat(linspace(1,0,overlap(1)).',[1 s0(2)]); % transition masks
mys = repmat(linspace(0,1,overlap(1)).',[1 s0(2)]);
mxw = repmat(linspace(1,0,overlap(2)),[st(1) 1]);
mxe = repmat(linspace(0,1,overlap(2)),[st(1) 1]);
colpict = zeros([st(1) s0(2) size(A,3) tiling(2)]); % preallocate
outpict = zeros([st size(A,3)]);
% assemble the frames into columns
f = 1;
for n = 1:tiling(2)
for m = 1:tiling(1)
if m == 1
sr = s0(1)-overlap(1);
ry = 1:sr;
colpict(ry,:,:,n) = A(1:sr,:,:,f);
elseif m == tiling(1)
sr = s0(1)-overlap(1);
ry = ry(end)+1:ry(end)+sr;
colpict(ry,:,:,n) = A(overlap(1)+1:end,:,:,f);
sr = s0(1)-2*overlap(1);
ry = ry(end)+1:ry(end)+sr;
colpict(ry,:,:,n) = A(overlap(1)+1:overlap(1)+sr,:,:,f);
% insert transition
if m ~= tiling(1)
ry = ry(end)+1:ry(end)+overlap(1);
colpict(ry,:,:,n) = A(end-overlap(1)+1:end,:,:,f).*myn ...
+ A(1:overlap(1),:,:,f+1).*mys;
f = f+1;
% assemble the columns into an image
for n = 1:tiling(2)
if n == 1
sr = s0(2)-overlap(2);
rx = 1:sr;
outpict(:,rx,:) = colpict(:,1:sr,:,n);
elseif n == tiling(2)
sr = s0(2)-overlap(2);
rx = rx(end)+1:rx(end)+sr;
outpict(:,rx,:) = colpict(:,overlap(2)+1:end,:,n);
sr = s0(2)-2*overlap(2);
rx = rx(end)+1:rx(end)+sr;
outpict(:,rx,:) = colpict(:,overlap(2)+1:overlap(2)+sr,:,n);
% insert transition
if n ~= tiling(2)
rx = rx(end)+1:rx(end)+overlap(2);
outpict(:,rx,:) = colpict(:,end-overlap(2)+1:end,:,n).*mxw ...
+ colpict(:,1:overlap(2),:,n+1).*mxe;
clear colpict
In this example, each supercolumn of the image is assembled from single-source blocks and transition blocks betwen them. This takes up a lot of space with indexing, but the interpolation needs to be applied to the minimum number of pixels. The interpolation itself is a simple linear transition implemented with multiplicative composition using precalculated masks.
The supercolumns are assembled into the final image in a similar fashion.
  15 Commenti
DGM il 18 Set 2021
I don't think the issue is with the interpolation. It looks like the data in each image has some offset. Normalizing the input images gets rid of the offsets. If that's the case, I don't know where the offset came from or what it means in this context, so I don't know if it's unwanted information.
A = cat(4,aa.hgtDep,cc.hgtDep,bb.hgtDep,dd.hgtDep);
outpict = mat2gray(imtile(A,[2 2],'direction','col'));
If the inputs are normalized, the question becomes which values should be used as references for denormalizing. If the inputs instead have their offsets removed somehow, the question becomes which tile is the calibrated one? Without figuring out where the offsets are coming from, I don't think I can answer either.
Paul Safier
Paul Safier il 22 Set 2021
@DGM Understood. Thanks for your help with this.

Accedi per commentare.

Più risposte (0)


Scopri di più su Read, Write, and Modify Image in Help Center e File Exchange




Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by