Hands-On Java Deep Learning for Computer Vision

Padding

One way to maybe increase the dimensions to 4 x 4 would be to introduce padding.

Padding by 1, for example, just adds one full rectangle of zeros, as shown in the following screenshot:

And by doing that, we increase the dimension of the input images from 6 to 8, so now we have 8 x 8 x 3, and by applying the formula we'll have, for example, 8-3+1=6. Now, notice how it was increased from 4 x 4 to 6 x 6; the 6 x 6 of the output is equal to the original input image, which is the blue color in the input matrix. In the literature, this phenomenon is called the Same Convolution, so the same convolution refers to the fact that the output convolve ...