The above images show (in order) the convolution of the cameraman image with D_x, D_y, the gradient magnitude and the application of a threshold of 0.25 to the grayscale image.
The above images show the original cameraman image and the blurred cameraman image with a 3x3 gaussian kernel with a sigma of 1.
The above images show the effects of 1.1 with a blurred cameraman image shown above.
The above images show the effects of the single convolution on the blurred image shown above.
The best rotation was at an angle of -2 degrees (the implementation searched from -20 to 20 degrees with a step of -2 degrees)
I took an image of Stonehenge and rotated it by 10 degrees. Then, the algorithm found the best rotation to be at an angle of 8 degrees.
I took an image of the leaning tower of Pisa and the algorithm found the best rotation to be at an angle of -4 degrees.
The above image fails because the tower is not in the middle. The algorithm finds the optimum angle to be at 0 degrees of rotation.