Mastering Computer Vision with TensorFlow 2.x

Padding

Padding is used to preserve the size of the feature map. With convolution, two issues can occur, and padding addresses both:

• The feature map's size shrinks with each convolution operation. For example, in the preceding diagram, a 7 x 7 feature map shrinks down to 5 x 5 due to convolution.
• Information at the edge is lost as the pixel on the edge is altered only once, whereas the pixel in the middle is altered many times by many convolution operations.

The following diagram shows a padding operation of size 1 being used on a 7 x 7 input image:

Note how padding conserves the dimension so that the output is the same size as the input. ...