Hands-On Artificial Intelligence for IoT

The convolution layer

This is the core building block of CNNs. It performs the mathematical operation similar to convolution (cross-correlation to be precise) on its input, normally a 3D image. It's defined by kernels (filters). The basic idea is that these filters stride through the entire image and extract specific features from the image.

Before going into further details, let's first see the convolution operation on a two-dimensional matrix for simplicity. The following diagram shows the operation when one pixel placed at position [2, 2] of a 5×5 2D image matrix is convolved with a 3×3 filter:

Convolution operation at a single pixel