Feature vector

A feature vector summarizes the (numerical) parametrizable properties of a pattern together in a vectorial manner. Various characteristic of the pattern features are the various dimensions of this vector. All of the possible feature vectors called the feature space. Feature vectors facilitate automatic classification, as they greatly reduce classifying properties (rather than a complete image must be viewed as only a vector of 10 numbers). Often they are used as input for a cluster analysis.

Examples

Speech recognition

In speech recognition, the energy of the speech signal is a commonly used feature. Furthermore, MFCCs, or which are based on linear prediction LPC, linear predictive coefficients (also: linear predictive coding) used, as well as the temporal variation of these quantities (first and second derivative with respect to time).

When the first 13 MFCCs, the corresponding derivatives, and the energy can be combined to form a feature vector, obtained 40 dimensions.

Prosodieerkennung

For automatic extraction of suprasegmental units in the Prosodieerkennung among others following basic features used:

  • The fundamental frequency F0 and the fundamental frequency variation
  • Various dimensions of the energy of the signal
  • Temporal extent of the speech signal, e.g., Pause lengths, Phonemlängen, etc

Image processing

  • Energy of the image
  • Fourier coefficients
  • Gray values

Text recognition and text analysis

  • Letters probability
  • Syllable probability
  • Word probability

Pattern classification

In the pattern classification pattern are parameterized based on their properties, the feature vectors are automatically classified. The better the features have been chosen and the more training material (ie, the larger the sample) is present, the more successful a classification. A larger dimension in the feature vectors in this context means a greater need for training material, including a larger amount of training and a greater duration. But it made ​​it even better classification rates, that is a better Klassifikatorqualität. A small number of dimensions this means a quicker training and a smaller sample, but lower quality.

Functions based on features as entries

Often, the basic features are offset by (weighted ) functions to aussagekräfigeren decision values ​​. These functions can calculate probability distributions or maximum likelihood form values ​​, percentages, ratios, a minimum, maximum or average.

Statistical classification Classifier (mathematics)
564868
de