Summary

In this chapter, we have presented the concept of a deep convolutional network, which is a generic architecture that can be employed in any visual processing task. The idea is based on hierarchical information management, aimed at extracting the features starting from low-level elements and moving forward until the high-level details that can be helpful to achieve specific goals.

The first topic was the concept of convolution and how it's applied in discrete and finite samples. We discussed the properties of standard convolution, before analyzing some important variants such as atrous (or dilated convolution), separable (and depthwise separable) convolution and, eventually, transpose convolution. All these methods can work with 1D, 2D, and 3D samples, even if the most diffused applications are based on bidimensional (not considering the channels) matrices representing static images. In the same section, we also discussed how pooling layers can be employed to reduce the dimensionality...