Project 1: Image Filtering and Hybrid Images
Example of a hybrid image.
Algorithm description
Below is the matlab code I wrote for my_imfilter.
##example code
padding_height = idivide (int32(size(filter,1)),int32(2),'floor');
padding_width = idivide (int32(size(filter,2)),int32(2),'floor');
pad_image = padarray(image,[double(padding_height) double(padding_width)],'symmetric');
output_image = zeros(size(image));
filter_height = size(filter,1);
filter_width = size(filter,2);
for k=1:size(image,3)
for i=1:size(image,1)
for j=1:size(image,2)
crop = pad_image(i:i+filter_height-1,j:j+filter_width-1,k);
C = crop.*filter;
output_image(i,j,k) = sum(C(:));
end
end
end
output = output_image;
- Padding: The padding choice I made was 'symmetric' (i.e. mirror reflection). the height/width of padding will vary upon shapes of the filter, determined by 'idivide' function. This will allow sides of the image to be filtered as well.
- Convolution: create an all zero matrix with same dimension as original picture. for each color dimension, rows and columns in original image, crop out same size of image as filter's, then perform element-wise multiplication for those two and set the pixel of interest to the convolution value in the all zero matrix. The matrix will eventually look as the filtered image.
Intermediate images
|
|
As we can see, when looking closely, the high frequency image (bike in this example) dominates the view. When viewing distance gets further away(or looking at a smaller scale), the low-frequency image becomes more perceivable.
On the other hand, the extra experiment of superman vs batman doesn't look as appealing. One reason might be that the alignment is not perfect. But I think the main reason is that amount of color in each image file is too small. Applying high-frequency filter on either image will leave only minimum outline image left, where the low-frequency image has very dominating yellow background. This is why the low-frequency image is always easier to identify from human perspective.