-6

I have a grayscale image, and need to increase its resolution. How can this be done in MATLAB? Would it mainly be done by multiplying the dimensions of the image for instance?

Charles
  • 50,943
  • 13
  • 104
  • 142
Simplicity
  • 47,404
  • 98
  • 256
  • 385

3 Answers3

2

You need to perform interpolation. There are many ways to do this. Use imresize (e.g. imgOut=imresize(img,scale,method);), or if you do not have the Image Processing Toolbox, consider the following code:

function imres = resizeim(I,outsize,interpalg)

if nargin<3 || isempty(interpalg),
    interpalg='cubic';
end

rows=outsize(1);
cols=outsize(2);

vscale = size(I,1) / rows;
hscale = size(I,2) / cols;

imgClass = class(I);
imres = interp2(double(I), (1:cols)*hscale + 0.5 * (1 - hscale), ...
                   (1:rows)'*vscale + 0.5 * (1 - vscale), ...
                   interpalg);
imres = cast(imres,imgClass);

Note: This is a rough start. You many need to perform pre-filtering, or other transformations. Also, this example only supports 2D (grayscale) images. For RGB, adapt this to process each color plane, or simple process each plane in a loop. Again, this is just an example.

Aside from edge handling, this gives the same results as imresize with anti-aliasing turned off (i.e. imresize(...,'Antialiasing',false)).

Regarding edge handling, see the documentation for interp2 for information on the extrapval parameter. The code gets ugly, but you can either patch the min/max elements in the interpolation points (interp2 inputs) to simply map exactly to the edges, or you can use NaN for extrapval, and post-process imres to replace the NaNs with its neighbor, etc. Note that simply interpolating at points such as linspace(1,size(I,1),rows) will not give the expected scale change.

chappjc
  • 30,359
  • 6
  • 75
  • 132
1

You can also perform sinc interpolation by Fourier transforming the image, zero-padding it, inverse Fourier transforming it, and taking the absolute value. 

im_rz = abs(ifft2(padarray(fft2(im),[row_pad, col_pad])))
Frederick
  • 1,271
  • 1
  • 10
  • 29
0

You can "hallucinate" high resolution details. See for example Glasner et al "Single image Super resolution" ICCV 2009.

An implementation of that algorithm can be found here

Shai
  • 111,146
  • 38
  • 238
  • 371