Vector quantization

The vector quantization is a method of compressing or identification of records.

The records are grouped into feature vectors. The idea of ​​the method is to assign these feature vectors that vector from a table, which is the observed feature vector most similar. Instead of storing all the information of the feature vector, only the index of this vector is most similar needs, see also database index.

The vector consists of two steps. In the first step, the training, the table ( codebook ) must be created with common feature vectors. In the second step, each of the codebook vector is determined by the smallest distance for other vectors. For transmitting data, only the index of the codebook vector is required, which may also be a vector when the code book is multidimensional. The corresponding decoder must have the same codebook and can then from the index generate an approximation of the original vector.

Another possible application is in the mapping of data records to particular patterns, as in the speech recognition. In this case, the distance between the feature vector and said codebook vector is used to decide whether the observed data set corresponding to a pattern.