Regarding the Gaussian filtering function imgaussfilt and the filter function fspecial.

I want to perform Gaussian filtering on an image, but the size of the filter will continue to increase as the program progresses, while the image is getting smaller and smaller. This inevitably leads to a situation where the image size is smaller than the filter size. So, I used these two functions to give a simple example to see if they can work properly.

Now that the execution has ended, I find that the results of the two functions are not the same. I would like to ask if these two functions have different ways of handling the issue of exceeding the matrix boundary? And which function would be more recommended for Gaussian filtering of images in image processing?

a = [1 2 1; 2 1 2; 1 2 1];
% imgaussfilt函数滤波对于边界像素的处理
b = imgaussfilt(a, 1.6);
disp(b);

    1.3271    1.3532    1.3271
    1.3532    1.3754    1.3532
    1.3271    1.3532    1.3271

% 使用same约束矩阵大小，对边界像素的处理
sigma = 1.6;
size = 9;
gaussian_kernel = fspecial('gaussian', size, sigma);
% disp(gaussian_kernel);
c = conv2(a, gaussian_kernel, 'same');
disp(c);

    0.4765    0.5545    0.4765
    0.5545    0.6450    0.5545
    0.4765    0.5545    0.4765

[Umar]

Hi 第四,

Please see my response to your posted questions.

Now that the execution has ended, I find that the results of the two functions are not the same.

Comments: The discrepancy in the results arises due to the different boundary handling methods employed by the imgaussfilt function which internally handles boundary pixels differently than manual convolution with the 'same' option. When using conv2 with the 'same' option, the output size matches the input size, leading to different boundary pixel treatment compared to imgaussfilt. However, if you want to align the results, consider adjusting the boundary handling method or padding strategy in manual convolution to match the behavior of imgaussfilt.

If these two functions have different ways of handling the issue of exceeding the matrix boundary?

imgaussfilt uses boundary padding techniques internally, while conv2 with the 'same' option constrains the output size to match the input size, affecting boundary handling differently.

And which function would be more recommended for Gaussian filtering of images in image processing?

imgaussfilt is often preferred due to its built-in boundary handling mechanisms, providing smoother results near the edges. However, if precise control over boundary handling is required, conv2 with appropriate padding strategies can be a better choice by considering the specific requirements of your image processing.

I hope this answers your question.

[第四]

感谢回答，考虑到之后的要求，可能会选择imgaussfilt函数，保证图像有更好的边缘

[Umar]

Hi 第四,

No problem, glad to help out. Sounds good to me.

[DGM]

@Umar is correct. The big advantage to imgaussfilt() is that it handles boundary regions sensibly and appropriately scales the kernel support by default.

FWIW, this is roughly how one would configure a fspecial()/imfilter() process to match the behavior of imgaussfilt():

% parameters
a = [1 2 1; 2 1 2; 1 2 1];
sigma = 1.6;

% imgaussfilt函数滤波对于边界像素的处理
b = imgaussfilt(a, sigma); % it's concise and convenient

% 使用same约束矩阵大小，对边界像素的处理
fsize = 2*ceil(2*sigma) + 1; % yes, that's 9 in this case.
gaussian_kernel = fspecial('gaussian', fsize, sigma); % filter with appropriate support
c = imfilter(a, gaussian_kernel,'replicate'); % specify the padding behavior

% just some negligible rounding error
immse(b,c)

ans = 2.1913e-32

Replicating the behavior with conv2() is more complicated, since the padding and cropping need to be done explicitly.