Wonnie Sim (ssim)

Hybrid image of Derek and his cat, Nutmeg

Algorithm

To apply Gaussian filter, image has to have an appropriate pad around it in order not to lose any boundary pixels after applying to filter. After setting up pad around the original image, take dot product to each block which has exactly the same size with filter. Laplacian filter is (original image - Gaussian filter applied image).

Result

As an image gets applied Gaussian filter, it loses distinct outlines and focuses more on colors, while an image with Laplacian filter focuses more outlines losing color information. When people see an object in close distance, they focuses more on its detailed shape becuase they can see the clear contour. However, with far distance, people cannot distinguish all the contours but try to find an object's outline with its color change. Thus, in hybrid image, high frequency image (which has clearer outline) is easily recognizeable with short distance while low frequency image (which has more color information) is easily recognizeable with far distance.

Example code

Below is a part of an example which I coded for this project.

%example code (filtering function)
for j = 1:size(image, 3)
    makeframe=padarray(image(:,:,j), [width, height], 'replicate');
    for i=1:size(image,1)*size(image,2)
            filtered(i)=sum(filtocol.*imtocol(:,i));
    end
    output(:,:,j)=reshape(filtered, size(image,1), []);
end

Result Table - given images

Result Table - new images

Gaussian Pyramid and Laplacian Pyramid

While working on above project, I became wondering the progress of losing contour and getting clearer contour. Thus I tried to construct Gaussian Pyramid by applying Gaussian filter several times and Laplacian Pyramid by deducting Gaussian filter applied images from the original image. The results are below.