How interp2 deal with edges on bicubic interpolation?

39 views (last 30 days)
Hi, I'm trying to make my own algorithm for 2d interpolation.
The upper left figure shows the data used for the interpolation. The upper right is the data interpolated with a linear method. The lower left is the result of matlab interp 2 cubic method and the lower right is an algorithm I made for cubic interpolation in 2d.
I wanted to know what interp2 does when you have data on the edge (for example the point where the X is). You need 2 points both sides to interpolate with the cubic method so what I did is interpolate linearly on the edge of the map. I'm on MATLAB 2016.
At the moment, my method uses separable convolution so I interpolate on rows first and then I interpolate on columns.
From what I understand from the code, interp2 extrapolates to be able to interpolate near edges, but I don't know which method is used to do so? Is it simply linear extrapolation?

Accepted Answer

Bruno Luong
Bruno Luong on 7 Jul 2021
Edited: Bruno Luong on 7 Jul 2021
The boundary handling is descriibed in the section Boundary Condition of this Cubic interpolation reference
Especially the equation (19) gives the value of the left outside pixel from the inside pixels
c(-1) = 3*f(x0) - 3*f(x1) + f(x2)
Likewise eqt (25) is for the right-side
c(N+1) = 3*f(x_N) - 3*f(x_{N-1}) + f(x_{N-2})
  12 Comments
Josh Meyer
Josh Meyer on 9 Jul 2021
Edited: Josh Meyer on 9 Jul 2021
It's important to note that the reference paper mentioned in the 2013 thread applied to interp2(..,'cubic') and interp1(..,'v5cubic') when it was originally posted. However, at that time interp1(..,'cubic') had different behavior and was the same as interp1(..,'pchip'), so the reference paper didn't apply there. In other words, interp2 and interp1 had similar options named 'cubic' that actually were different methods, and I believe this is the source of your confusion.
As of R2020b the behavior of interp1(..,'cubic') is the same as interp1(..,'v5cubic'), so the reference paper now applies for interp1's cubic option as well. The details of this change are at the bottom of the interp1 page.
Bruno Luong
Bruno Luong on 9 Jul 2021
interp(...'pchip') is NOT the same as interp1(..., 'v5cubic'). The later uses convolution formula on the paper. And 'pchip' preserves data monotony and it is a non-linear method.
So 'cubic' in R2013 cannot match both IMO.
Anyway it is not big deal, I'm only interested in what happens now.

Sign in to comment.

More Answers (2)

Kelly Kearney
Kelly Kearney on 7 Jul 2021
interp2 does not extrapolate. By default, it sets any points outside the data bounds to NaN; you can alternatively choose a different scalar value to assign to all extrapolated points. Alternatively, you could use griddedInterpolant, which allows you to explicitly assign an algorithm to the extrapolated points:
[x,y,z] = peaks(5);
[xi,yi] = meshgrid(linspace(-4,4,50));
zi1 = interp2(x,y,z,xi,yi,'cubic');
F = griddedInterpolant(x',y',z,'cubic','linear');
zi2 = F(xi',yi');
subplot(2,1,1);
hi = imagesc(xi(1,:),yi(:,1),zi1, 'AlphaData', ~isnan(zi1), 'alphadatamapping', 'scaled');
hold on;
scatter(x(:), y(:), [], z(:), 'filled', 'markeredgecolor', 'w');
title('interp2: cubic')
subplot(2,1,2);
hi = imagesc(xi(1,:),yi(:,1),zi2, 'AlphaData', ~isnan(zi2), 'alphadatamapping', 'scaled');
hold on;
scatter(x(:), y(:), [], z(:), 'filled', 'markeredgecolor', 'w');
title('griddedInterpolant: cubic, linear')
  1 Comment
Jérémy Talbot-Pâquet
Jérémy Talbot-Pâquet on 7 Jul 2021
Edited: Jérémy Talbot-Pâquet on 7 Jul 2021
Thank you for your answer. Sorry if I wasn't clear, I meant that somewhere in interp2 for the cubic method, the array is expanded outside the domain to be able to interpolate on the edges since you need 2 points both sides. But you're right, interp2 does not return an extrapolated array unless you give an extrapval argument.
I haven't thought of GriddedInterpolant it could help me to expand the array to more than 1 point. Thank you

Sign in to comment.


Matt J
Matt J on 7 Jul 2021
Edited: Matt J on 7 Jul 2021
Based on a little reverse engineering, I believe I have discovered that the 'cubic' method of interp1 uses quadratic extrapolation to pad 1 element at the beginning and end of the array. I think we can be pretty confident that the obvious generalization to the 2D case is used for interp2.
To verify this, the test below checks that the extrapolation rule leads to the same interpolated values regardless of whether the signal starts at the beginning of the array or in the middle:
v1=[rand(1,5) 0 0 0 0 0],
v1 = 1×10
0.1348 0.1215 0.0821 0.6977 0.1237 0 0 0 0 0
t1=0:numel(v1)-1;
F1=@(x) interp1(t1,v1,x,'cubic');
p=polyfit(0:2, v1(1:3),2);
extrap=polyval(p,-1), %quadratic extrapolation
extrap = 0.1222
v2=circshift(v1,5); %shift v1 and insert extrapolated value
v2(5)=extrap,
v2 = 1×10
0 0 0 0 0.1222 0.1348 0.1215 0.0821 0.6977 0.1237
t2=t1-5;
F2=@(x) interp1(t2,v2,x,'cubic');
t=linspace(0,1,30);
plot(t,F1(t),'-',t,F2(t),'o')
xlabel t
ylabel 'Interpolated Values'
legend('v1(t1)','v2(t2)')

Categories

Find more on Interpolation in Help Center and File Exchange

Products


Release

R2016a

Community Treasure Hunt

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

Start Hunting!