how can i fill the region inside the leaf

I was using the k-mean segmentation with a* space in L*a*b* color space for removing the background image but unfortunatly some of the diseased part was equal to center of the bakground, i want to add these small diseased spot areas to the foreground, can anyone help me with what should i do to solve it.

Answer 1

[Walter Roberson]

rgb = reshape(YourImage, [], 3);
idx = kmeans(rgb, NumberOfClusters) ;
labeled = reshape(idx, size(YourImage(:, :, 1)));
isleaf = labeled == ClusterNumberForLeaf;
filledleaf = bwfill(isleaf, 'holes');

You need to figure out how to decide which cluster number is the leaf; it will not be the same every time.

[Alaa Eldesoky]

filledleaf = bwfill(isleaf, 'holes');

it doesn't give agood result too

the k-mean with a* space code run with me with all images and give a good results in segmentation, only just in the images that have a dead cell issue in the leaf so it become equal to the value of the soil and i know the programe have to detect the disease before this happening, but i was thinking just if it happening, how could i segment it.

I'm thinking in removing the pixel areas that is quite small from the backgroung image and add it to the foreground image but i don't know the right way to do that.

[Walter Roberson]

~bwareafilt(~isleaf, [1 SIZE] )

would fill any holes up to area SIZE

Answer 2

[Image Analyst]

kmeans() is not usually a good way for color segmentation (my demo is attached) regardless if you're doing it in RGB or any other color space. Why not? Well, for one illustration, imagine you want the area fraction of leaf in the field of view with any amount from 0% to 100%. Now imagine you're looking at an image with virtually no leaf in there -- just dirt. If you tell kmeans to find 2 clusters, it will find two clusters that will be different shades of dirt. None of them will be green leaf color. So now, after segmentation by kmeans you're still going to have to examine the clusters to see which one, if any, are green. What if you have a small amount of green, like 5% and lots of dirt? It still might find clusters that are the two dominant shades of brown and you'd totally miss the green leaf pixels. It's better to use fixed thresholds.

You'd be better off with a segmentation that thresholded on the hue channel using fixed threshold of hues known to be green. Use the Color Thresholder App on the Apps tab of the tool ribbon to threshold in HSV color space. Then find small blobs in the background mask with bwareaopen or bwareafilt. Then OR the small blobs mask in with the leaf mask to fill in the small holes only.

hsvImage = rgb2hsv(rgbImage);
hImage = hsvImage(:, :, 1);
leafMask = hImage > hue1 & hImage < hue2;
backgroundMask = ~leafMask;
% Get blobs in the background up to some max size, like 100 pixels.
maxAcceptableSize = 100;
smallBlobs = bwareafilt(backgroundMask, [1, maxAcceptableSize]);
% We want to add/OR those small blobs into the leaf mask so that
% they will be considered as leaf, not background.
leafMask = leafMask | smallBlobs;

[Alaa Eldesoky]

after alot of attempt I used this code for segmentation with selected initial centroid and it seems to give more good result than colorthreshold as i used it first to make a background subtraction but it doesn't give me this results

just i have the problem of background subtraction in the images like i show recently

% Read the image
[filename, pathname] = uigetfile({'*.*';'*.bmp';'*.jpg';'*.gif';'*.JPEG'}, 'Pick a Leaf Image File');
Image = imread([pathname,filename]);
%% resize the image
   imgresized = imresize(Image,[1400 1200],'bicubic');
%% split Channel and applying median filter on it
   redChannel = imgresized(:, :, 1);
   greenChannel = imgresized(:, :, 2);
   blueChannel = imgresized(:, :, 3);
   redMF = medfilt2(redChannel, [3 3]);
   greenMF = medfilt2(greenChannel, [3 3]);
   blueMF = medfilt2(blueChannel, [3 3]);
   imgMF = cat(3, redMF, greenMF, blueMF);    
%% applying K-mean clustering with a* Channel in Lab color space
   cform = makecform('srgb2lab');
    % Apply the colorform
   labimg = applycform(imgMF,cform);
    % Classify the colors in IPCA using K means clustering.
    achannel = double(labimg(:,:,2));
    nrows = size(achannel,1);
    ncols = size(achannel,2);
    a = reshape(achannel,nrows*ncols,1);
    nClusters = 2;    
%% Specify the centroid to automatically segment the image for the whole leaf and background
    intialcentroid= [105 ; 135];
    [idx]= kmeans(a,[],'start',intialcentroid);
    % Label every pixel in tha image using results from K means
    pixel_labels = reshape(idx,nrows,ncols);
    % Create a blank cell array to store the results of clustering
    segmented_images = cell(1,3);
    % Create RGB label using pixel_labels
    rgb_label = repmat(pixel_labels,[1,1,3]);
    for k = 1:nClusters
    colors = imgMF;
    colors(rgb_label ~= k) = 0;
    segmented_images{k} = colors;
    end
%% imclose and show the image
    SEa = strel('disk',4);
    segmentedimg1 = imclose(segmented_images{1},SEa);
    segmentedimg2 = imclose(segmented_images{2},SEa);
  figure, subplot(1,3,1);imshow(imgresized);axis tight equal;title('image Mobile Data');
  subplot(1,3,2);imshow(segmentedimg1);title('Cluster 1 Leaf'); 
  subplot(1,3,3);imshow(segmentedimg2);title('Cluster 2 Background');  

[Image Analyst]

Yeah, it's bad. Hopefully now you'll understand why I said what I said - kmeans is no good for color segmentation and you should use fixed thresholds. Not sure what went wrong when you tried the Color Thresholder.

An alternative is to use discriminant classification. I'm attaching a demo.

[Alaa Eldesoky]

oh! okay, I get what you mean, I'll try again the HSV thresholding