sjtrny / dark-channel-haze-removal Goto Github PK

Hello, I am honored to read your paper. I would like to ask if this code is the original code of He Keming?

@sjtrny

Hi Stephen, very nice code. I am trying to reproduce this method in R. I have a small question I came across while porting: in get_atmosphere.m

n_search_pixels = floor(n_pixels * 0.01);

In the reference paper (http://research.microsoft.com/en-us/um/people/jiansun/papers/Dehaze_CVPR2009.pdf) it is stated

0.1% bright-est pixels in the dark channel. These pixels are most haze-opaque (bounded by yellow lines in Figure6(b)).

It doesn't make much difference (I would say) at the end since you take the average of the accumulated pixels, but I was wondering why you used 1% instead of 0.1%.

Many thanks in advance.

In your get_atmoshpere.m code, you first select top 1% percent brightest pixels in the dark channel. Then, the average intensity of these pixels in the original input hazy images is regarded as the atmospheric light. However, this does not match the original paper (Single Image Haze Removal Using Dark Channel Prior) that illustrates the atmospheric light is considered as the pixels with highest intensity (mean intensity of each channel) in the input hazy image I are selected the atmosphere light.

Therefore, I provide my revised code based on your given code.
function atmosphere = get_atmosphere(image, dark_channel)

[m, n, ~] = size(image);
n_pixels = m * n;

n_search_pixels = floor(n_pixels * 0.001);

dark_vec = reshape(dark_channel, n_pixels, 1);

image_vec = reshape(image, n_pixels, 3);

[~, indices] = sort(dark_vec, 'descend');

% accumulator = zeros(1, 3);

% for k = 1 : n_search_pixels
% accumulator = accumulator + image_vec(indices(k),:);
% end
%
% atmosphere = accumulator / n_search_pixels;

atmosphere_search_pixels = zeros(n_search_pixels,3);
for k=1 : n_search_pixels
atmosphere_search_pixels(k,:) = image_vec(indices(k),:);
end

[~, top_intensity_indices] = sort(mean(atmosphere_search_pixels,2), 'descend');
atmosphere = atmosphere_search_pixels(top_intensity_indices(1),:);

end

Hi @sjtrny
I was looking into get Laplacian and I was testing it with a very simple image (resize of the original forest image to 10% of the original size). I see that the win_inds array becomes:

Empty matrix: 1-by-0

since n_min becomes bigger than n_max.

(When k is around 8010 just to give you a reference.)
The emptiness of the matrix has problems that hit then num_neigh that becomes 0 and win_var becomes NaN.
Is this behavior on purpose? I quite don't get it. Can you, please, give me some details on this? Since of course R, to which I am porting the code and idea, complains much more than matlab already when the indexes of the win_inds don't match anymore the dimensions.

Many thanks in advance.