One class of artificial neural networks that are particularly good for processing image information (localization of image pixels, etc.) are convolutional neural networks (CNNs).
Basically, all networks consist of a part that is used for feature discovery. Here, covolutional layers, in which discrete convolutions are performed, alternate with a pooling layer. The most widely used pooling method is max-pooling. Here, the neurons of the covolution layer are sampled with an nxn square and only the most active neuron is taken over in each case. In the following figure this method is exemplarily shown with a 2×2 filter.
The feature finding part is usually followed by a classification through one or more fully-connected layers. These pass location-independent object information to an output layer. The number of the last layers corresponds to the number of specific classes.