It's the mean difference between a pixel and it's neighbors. So you'd convolve with an array [-1 1] in each of the 8 directions. Since each direction is unique, the convolution is separable and you can just add up all the kernels. So you'd get a 3x3 window with 8 -1's all the way around the perimeter, and 8 1'd in the middle. In other words, the 3x3 matrix would have -1 all the way around except a value of 8 in the center:
kernel = -1 * ones(3);
kernel(2,2) = 8; % Now kernel = [-1,-1,-1; -1,8,-1; -1,-1,-1]
output = conv2(double(inputImage), kernel, 'same');
