Feature vector
A feature vector summarizes the (numerically) parameterizable properties of a pattern in a vectorial manner. Different features characteristic of the pattern form the different dimensions of this vector. The totality of the possible feature vectors is called the feature space . Feature vectors facilitate automatic classification , as they greatly reduce the properties to be classified (for example, instead of a complete image, only a vector of 10 numbers has to be considered). They often serve as input for a cluster analysis .
Examples
voice recognition
In speech recognition , the energy of the speech signal is a frequently used characteristic. Furthermore, MFCCs or the LPCs based on linear prediction , linear predictive coefficients (also: linear predictive coding ) are used, as well as the temporal change of these variables (first and second derivative according to time).
If the first 13 MFCCs, the associated derivatives and the energy are combined to form a feature vector, 40 dimensions are obtained.
Prosody recognition
For the automatic extraction of suprasegmental units, prosody detection u. a. the following basic features are used:
- The basic frequency F0 or the basic frequency curve
- different measures of the energy of the signal
- temporal measures of the speech signal, e.g. B. pause lengths, phoneme lengths etc.
Image processing
- Energy of the image
- Fourier coefficients
- Gray values
Text recognition and text analysis
- Letter probability
- Syllable probability
- Word probability
Pattern classification
In the pattern classification , patterns are automatically classified based on their parameterizable properties, the feature vectors. The better the characteristics were chosen and the more training material (i.e. the larger the sample) is available, the better a classification succeeds. A larger dimension in the feature vectors means a greater need for training material, that is, a greater training effort and a longer training duration. But you also get better classification rates, i.e. better classifier quality. A small number of dimensions means faster training and a smaller sample, but also lower quality.
Functions based on basic features as entries
The basic features are often calculated using (weighted) functions to form more meaningful decision values. These functions can compute probability distributions or form maximum likelihood values, percentage values, ratio values, a minimum, maximum or an average.