How can I detect the defect in this image of a semiconductor?

Question

So I have this image of a semiconductor wafer and it has a defect in it which I need to detect using Matlab. I should detect only it and not its background. I need to also measure its perimeter and area.

So far I have this code where I convert the original image to a binary image and then use dilation on it and then try to get its contour. When getting the perimeter and the area, I receive the outcome not only for the defect but for the rest of the image which is not what I want. How can I extract the defect only so I can get only its area and parameter.

The image:

fig3 = imread('figure3.png');
imshow(fig3);

title('Original image', 'FontSize', 18);
%Gray image
fig3Gray = rgb2gray(fig3);



%Binary image
BW3 = imbinarize(fig3Gray,0.5);
imshow(BW3);
title('Binary image', 'FontSize', 18);

se3 = strel('square',5);

%Dilation image
dilated3 = imdilate(BW3,sr);
imshow(dilated3);
title('Dilated image', 'FontSize', 18);

minus3 = ~(BW3-dilated3);
imshow(minus3);
title('Contour image', 'FontSize', 18);
imshowpair(minus3,BW3,'montage');

%Perimeter and Area calculation
Coon3 = bwconncomp(~BW3)
ANS3 = length(Coon3.PixelIdxList{1});
OUTPUT3 = regionprops(Coon3,'Perimeter','Area');
P3 = OUTPUT3.Perimeter
Area3 = OUTPUT3.Area

Answer 1

Let's start by reading the image and converting it to binary. Note that I have lowered the threshold to eliminate unwanted details.

clear; close all; clc

fig3 = imread('XEQ59.png');
imshow(fig3);

title('Original image', 'FontSize', 18);
%Gray image
fig3Gray = rgb2gray(fig3);
%Binary image
BW3 = imbinarize(fig3Gray, 0.2); % lowered threshold
figure; imshow(BW3)
title('binary image')

---

Now we move on to find the coordinate of the defect. To do this, we define a structuring element that defines our desired shape se.

We are looking for the parts in the image that match with se. For a given coordinate to match, the surrounding area must exactly be se.
Note that gray values here are ignored, they can be either white or black.

se is manually defined where 1 represents white, -1 represents black, and 0 represents ignored pixels.

% hit-miss
se = [1, 1, -1*ones(1,5), ones(1, 3); ...
      ones(6,1), -1*ones(6), zeros(6,2), ones(6,1); ...
      ones(3,2), zeros(3,1), -1*ones(3,6), ones(3,1)];
figure; imshow(uint8(255/2*(se+1)), 'InitialMagnification', 3000)
title('structuring element')

---

Applying hit-miss operation to find the position of the defect:

pos = bwhitmiss(BW3, se);
figure; imshow(pos)
title('position of defect')
input('Press enter to continue...')

---

Now that we have the position, we grow that particular pixel position until it grows no more, in order to obtain defect.

% get the defect
close all; clc
def = pos;
last_def = zeros(size(def));
while ~isequal(def, last_def)
    last_def = def;
    def = ~BW3 & imdilate(def, ones(3));
    imshow(def)
    title('defect')
    pause(0.1)
end

---

Calculating the area and the perimeter:

% area
area = sum(def(:))

% perimeter
vert = imdilate(def, [1; 1; 1]) - def;
horz = imdilate(def, [1 1 1]) - def;
perimeter = sum(vert(:)) + sum(horz(:))

area =
   102
perimeter =
    54

Answer 2

This question is much more difficult then your previous question.

• Following solution uses an iterative approach (two iterations).
• Includes heuristic about the difference from the cluster to it's neighbors.
• Includes a heuristic that the cluster can't be too tall or too long.

Please read the comments:

clear

fig3 = imread('figure3.png');
fig3Gray = rgb2gray(fig3);
fig3Gray = im2double(fig3Gray); %Convert from uint8 to double (MATLAB math works in double).

%figure;imshow(fig3Gray);

%Apply median filter with large radius.
Med = medfilt2(fig3Gray, [51, 51], 'symmetric');
%figure;imshow(Med);

D = abs(fig3Gray - Med);
%figure;imshow(D);impixelinfo

BW = imbinarize(D, 0.3);
%figure;imshow(BW);impixelinfo

Coon = bwconncomp(BW);

fig3GrayMasked = fig3Gray;

%Cover the tall clusters and the long clusters.
for i = 1:length(Coon)
    C = Coon.PixelIdxList{i}; %Cluster coordinates.
    [Y, X] = ind2sub(size(fig3Gray), C); %Convert to x,y indices.
    is_tall = (max(Y) - min(Y)) > 50; %true if cluster is tall.
    is_wide = (max(X) - min(X)) > 50; %true if cluster is wide.

    %Replace tall and long clusters by pixels from median image.
    if ((is_tall) || (is_wide))
        fig3GrayMasked(C) = Med(C);
    end
end

%figure;imshow(fig3GrayMasked);impixelinfo

%Second iteration: search largest cluster on fig3GrayMasked image.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Med = medfilt2(fig3GrayMasked, [51, 51], 'symmetric');

D = abs(fig3GrayMasked - Med);
%figure;imshow(D);impixelinfo

BW = imbinarize(D, 0.3);
%figure;imshow(BW);impixelinfo

Coon = bwconncomp(BW);

%Find index of largest cluster in list of clusters Coon.PixelIdxList
[~, i] = max(cellfun(@numel, Coon.PixelIdxList));

%Get the indices of the largest cluster
C = Coon.PixelIdxList{i};
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%Paint cluster in yellow color (just for fun).
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
BW = zeros(size(BW), 'logical');
BW(C) = 1;
Y = im2uint8(cat(3, ones(size(BW)), ones(size(BW)), zeros(size(BW))));
fig3(cat(3, BW, BW, BW)) = Y(cat(3, BW, BW, BW));
figure;imshow(fig3)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Result: