Canny Edge Detection Algorithm problem?

Hello Dear Experts,

Here is what I have written as an implementation of the Canny Edge detection algorithm but I get non accurate results. Please try to help me, what can be the reason, where did I do a mistake?

function [ newImg ] = edgeDetect( img )

% Step 0: Initialization, smoothing the matrix to eliminate the noise:

[imgRows imgCols] = size(img);

% Gaussian Filter:
GaussianFilter = fspecial('gaussian', 5, 1);
%T = T/sum(sum(T)); % Normalizing the filter
SmoothMat = conv2(img,GaussianFilter,'same'); % Smoothing

epsilon = exp(-10);
newImg = zeros(size(img));
HiT = 0.15;
LoT = 0.07;

% Step 1: Calculating the derivative of the smoothed image using convolution:

dx = conv2(SmoothMat,[-1 1],'same');
dy = conv2(SmoothMat,[-1 1]','same');

% Substep: There can be dx(i,j) = 0 and dy(i,j)/dx(i,j) will be +/-inf to
% avoid it we will add an epsilon to eliminate the +/-inf problem:

idx = dx == 0;
dx(idx) = epsilon;

% Step 2: Calculating the gradient and the angle:

GradVec = sqrt(dx.*dx + dy.*dy);
TanTheta = dy./dx;
% Avoiding the edges of the matrix - first and last rows and columns
TempGradVec = padarray(GradVec,[1 1]);
TempTanTheta = padarray(TanTheta,[1 1]);

% Step 3: Checking direction condition and gradient value along with NMS

for i = 2:imgRows+1
    for j = 2:imgCols+1

          if ((TempTanTheta(i,j) >= -0.4142) && (TempTanTheta(i,j) <= 0.4142))

              TempGradVec(i,j) = isMax(TempGradVec(i,j-1),TempGradVec(i,j),TempGradVec(i,j+1));

          elseif ((TempTanTheta(i,j) > 0.4142) && (TempTanTheta(i,j) <= 2.4142))

              TempGradVec(i,j) = isMax(TempGradVec(i+1,j-1),TempGradVec(i,j),TempGradVec(i-1,j+1));

          elseif (abs(TempTanTheta(i,j)) > 2.4142)

              TempGradVec(i,j) = isMax(TempGradVec(i-1,j),TempGradVec(i,j),TempGradVec(i+1,j));

          elseif ((TempTanTheta(i,j) >= -2.4142) && (TempTanTheta(i,j) < -0.4142))

              TempGradVec(i,j) = isMax(TempGradVec(i-1,j-1),TempGradVec(i,j),TempGradVec(i+1,j+1));

          end
      end
  end

% Step 4: The hysteresis part, using the predefined threshold values

for i = 2:imgRows-1
    for j = 2:imgCols-1
        if TempGradVec(i,j) > HiT
            TempGradVec(i,j) = 1;
        elseif TempGradVec(i,j) < LoT
            TempGradVec(i,j) = 0;
        elseif TempGradVec(i,j-1) == 1 ||...
                TempGradVec(i-1,j) == 1 ||...
                TempGradVec(i-1,j+1) == 1 ||...
                TempGradVec(i,j-1) == 1 ||...
                TempGradVec(i,j+1) == 1 ||...
                TempGradVec(i+1,j-1) == 1 ||...
                TempGradVec(i+1,j) == 1 ||...
                TempGradVec(i+1,j+1) == 1
            TempGradVec(i,j)=1; 
            TempGradVec(i,j)=1;
        else
            TempGradVec(i,j)=0;
        end
    end
end

newImg = TempGradVec(2:end-1,2:end-1);      

end

function [num] = isMax(p1,p2,p3)
 %if p2 is the biggest
if p2>=p3 && p2>=p1
    num=p2;
else
    num=0;
end
end

Answer 1

[Image Analyst]

No - I'm not into reinventing wheels for basic building block functions like this. Any reason why you don't just use the edge() function that's built in to MATLAB's Image Processing Toolbox? It does a Canny filter.

[Steve]

I need to write it myself as part of some big project.

I know there is such function but I need your professional help to understand what is wrong.

There are a lot of experts here who help a lot including you.

Answer 2

[Walter Roberson]

Consider your lines

if TempGradVec(i,j) > HiT
          TempGradVec(i,j) = 1;
      elseif TempGradVec(i,j) < LoT
          TempGradVec(i,j) = 0;

Those do convert (some) floating point results to exact 0's and 1's, and store those exact values back into TempGradVec over-top of the previous floating point values. But if the floating point value does not happen to be in that particular range, you go on to the test

elseif TempGradVec(i,j-1) == 1 ||...
              TempGradVec(i-1,j) == 1 ||...
              TempGradVec(i-1,j+1) == 1 ||...
              TempGradVec(i,j-1) == 1 ||...
              TempGradVec(i,j+1) == 1 ||...
              TempGradVec(i+1,j-1) == 1 ||...
              TempGradVec(i+1,j) == 1 ||...
              TempGradVec(i+1,j+1) == 1

The first 4 of those tests are looking at locations you have already converted, so it is safe to do exact comparisons to 1 on those locations. But starting from TempGradVec(i,j+1) you are examining a location that you have not already converted and so are looking at raw floating point values. Is it plausible, really, that in those unconverted sections of the matrix, any of the results will exactly equal 1 down to 53 bit binary floating point precision ?