Machine Learning for Developers

Kernels and convolutions

When solving practical problems in the discrete domain, we normally have 2D functions of finite dimensions (which could be an image, for example) that we want to filter through another image. The discipline of filter development studies the effects of different kinds of filters when applied via convolution to a variety of classes. The most common types of function applied are of two to five elements per dimension, and of 0 value on the remaining elements. These little matrices, representing filtering functions, are called kernels.

The convolution operation starts with the first element of an n-dimensional matrix (usually a 2D matrix representing an image) with all the elements of a kernel, applying the center element ...