Representation background

Representation background

Within PRTools representation is the process that converts objects given as raw data (e.g. images, time signals, spectra, possibly of varying sizes) into a vector space of a fixed, finite dimensionality. The raw data should be given as a dataset or a datafile. The conversion is done by fixed, user defined mapping, or by a mapping that is optimized for some data, a so-called trainable mapping.

The following representations of entire objects will be treated more extensively in separate subsections

• Features measured in images or in 1-D objects like spectra that may be considered as images.
• Pixels or signal samples used as vector representation of the entire object. This is different from the representation of pixels as objects, see below.
• Kernels based on inner products in an already given vector space.
• Dissimilarities based on some distance measure between objects.

There is a separate subsection of the representation of pixels in images by multiple images, spectra (hyperspectral images) or pixel neighborhoods. This is discussed separately.

Many routines in these sections are written for 2-dimensional images but can be applied to 1-dimensional images as well. To protect the user against misuse an error is generated in case a dataset does not contain object images. If these were not automatically created by PRTools a dataset A may be converted by

    a = setfeatsize(a,[1 getfeatsize(a)]) 

This only makes sense if the features in the dataset a are samples of some continuous measurement like a time signal, a spectrum or a histogram.

Representation background