Deep-Learning-with-PyTorch

Convolutions in action
201
0
output
0
input
5
5
10
10
15
15
20
20
25
25
30
0 5 10 15 20 25 30
30
0 5 10 15 20 25 30
Figure 8.4
Our bird, this time blurred thanks to a constant convolution kernel
Working out the weighted sum for an arbitrary pixel in position 2,2, as we did earlier
for the generic convolution kernel, we get
o22 = i13 - i11 +
i23 - i21 +
i33 - i31
which performs the difference of all pixels on the right of i22 minus the pixels on the
left of i22. If the kernel is applied on a vertical boundary between two adjacent regions
of different intensity, o22 will have a high value. If the kernel is applied on a region of
uniform intensity, o22 will be zero. It’s an edge-detection kernel: the kernel highlights the
vertical edge between two horizontally adjacent regions.
Applying the convolution kernel to our image, we see the result shown in figure
8.5. As expected, the convolution kernel enhances the vertical edges. We could build
0
output
0
input
5
5
10
10
15
15
20
20
25
25
30
0 5 10 15 20 25 30
30
0 5 10 15 20 25 30
Figure 8.5
Vertical edges throughout our bird, courtesy of a handcrafted convolution kernel