Applying a Gaussian Filter using no built-in MatLab functions

Question

I'm new to Matlab. I'm really just getting started into the actual computer science classes in my major. So, please keep that in mind. The goal is to create and apply a gaussian filter to this specific image using no built-in functions. So far, I have this bit of code to create a kernel. We were playing around with different sigma values, and then produced a visualization of the kernel.

    f = imread( 'input.png');

    sig = 5;
    hw = floor (2.5 * sig - .5);

    w = zeros(hw*2+1, hw*2+1);

    for r = 1:size(w,1)
        for c = 1:size(w,2)
            w(r,c) = exp(-1 * ((c - (hw+1))^2 + (r-hw)^2) / (2 * sig^2));
        end
    end

    imagesc(w);
    colormap jet;

My problem comes when actually applying it. I'm really not sure on what to do. He gave us the following code as a guide, but I'm still stuck.

    for r = 1:R
        for c = 1:C
            for r1 = 
                for c1 = 
                   temp = temp + f() + w();
                end
            end
        end
    end

If anyone could point me in the right direction, it would be greatly appreciated. Thank you.

What do you consider a "built-in"? Technically `floor`, `imread`, `zeros`, etc. are all builtins — Suever, Feb 17 '17 at 01:34
Sorry, I should've been more specific. He just says, "Do not use any built-in functions for building Gaussian filter nor performing convolution." We can't use any filter functions, convolution functions, etc. It's hard to explain because I don't have much Matlab knowledge. We have only learned basic functions, so that's all we're allowed to use. — user7578205, Feb 17 '17 at 02:34

Sukuya · Accepted Answer · 2017-02-17T07:41:07.010

Assuming it is a tutorial problem given to you in the class, I will give you hints to proceed.

Outer two loops should be used to traverse over the entire image. You need to put proper R and C depending upon the size of kernel you are using. (You have to avoid the boundary such that your kernel doesn't go outside the image.)
Before the beginning of inner two loops, you must reset your temp to zero, r1 and c1 must vary from 1 to size of the kernel.
f() and w() should be multiplied and added to the temp to effectively do convolution. Now your task is to work out the appropriate indices used for f and w.
And you need to update the current pixel value with temp at appropriate place.

I hope this makes sense.

Cheers,

score 1 · Answer 2 · edited Feb 17 '17 at 07:43

If you're willing to use a Fast Fourier Transform, everything becomes easy:

Fourier transform your image (component per component)
Multiply every transformed component element-by-element with a matrix containing your filter in frequency space.
Fourier transform back, and drop the small complex part your filter introduced.

Of course, if you consider fft a forbidden builtin, this doesn't qualify.

Applying a Gaussian Filter using no built-in MatLab functions

2 Answers2