Clarification on linear vs circular convolution

I have a simple program to satisfy myself that I understand convolution. This program takes an input image, calculates the 2D convolution using fft2, calculates the 2D convolution using conv2, then compares the results.

The code is below:

%%Image input
source_image_path = 'C:\...\cameraman.tif'; 
myFilter = fspecial('gaussian', [256 256], 4);

% load source image
myImage = im2double(imread(source_image_path)); 
if ( size(size(myImage)) == [1 3])
    myImage = rgb2gray(myImage);
end

subplot(2,2,1); % 2 rows, 3 columns, grid position 1
imshow(myImage, []);
axis on;
title('Source', 'FontSize', 15);

conv = fftshift(ifft2((fft2(myImage) .* fft2(myFilter)))) ;
subplot(2,2,2);
imshow(conv, []);
title('fft conv result', 'FontSize', 15);
axis on;

second_conv = conv2(myImage, myFilter, 'same') ;
subplot(2,2,3);
imshow(second_conv, []);
title('conv2 result', 'FontSize', 15);
axis on;

subplot(2,2,4)
difference = imshowpair(conv, second_conv, 'diff'); 
title('difference', 'FontSize', 15); 
axis on;

Testing this on cameraman.tif gives the following result:

So as expected, barring some fft errors near the edges. However, if I test on a different image, of tubuls, I get this:

Here the convolutions do not match. I went away and did some reading about linear vs circular convolution, and how to get these results to match using padding. I padded my code, and now my result is this:

So padding the code has successfully matched the linear and circular convolution methods.

My questions are therefore:

What is different about cameraman.tif and tubuls.tif that causes them to have different results in my unpadded code? I think this is something to do with Matlab assuming a periodicity in the tupuls image; why are the two images not treated the same?

Finally, which is the 'right' way to do this? I notice that retesting cameraman.tif on the padded code removes the edge defects. Should I always include padding when performing convolution using fft2? If so, what is the point of the linear convolution, and when might people use it?

[Rayyan Khan]

Did you get the answer of your question? please let me know thanks.