Method for vector quantization of a feature vector
View Patent ↗The invention relates to a method for the vector quantization of a feature vector, in particular with respect to a data compression of a signal to be transmitted or to be stored, particularly a voice signal or a video signal, wherein at least one codebook from a plurality of codebook vectors is searched for a code vector representing the feature vector. During the search, a sequence of codebook vectors is examined for the suitability thereof to represent the feature vector. In the course of the search for the code vector, a set of neighboring vectors is dedicated to at least one of the codebook vectors potentially to be examined, preferably prior to the search. The search for the code vector includes at least part of the neighboring vectors.
1. A method for vector quantization of a feature vector, comprising:
searching at least one codebook comprising multiple codebook vectors to find a code vector representing a feature vector;
during the search, testing a sequence of codebook vectors to determine their suitability to represent the feature vector; and
for at least one of the codebook vectors that can be tested during the search for the code vector representing the feature vector, determining a set of neighboring vectors before the beginning of the search and including in the search for the code vector representing the feature vector at least some of the neighboring vectors; and
determining the set of neighboring vectors for a codebook vector using a neighboring vector table that assigns a number of nearest neighboring vectors to each codebook vector from among the multiple codebook vectors in the at least one codebook, in which, starting with the codebook vector whose set of neighboring vectors is to be determined and which is entered as a start vector in the set of neighboring vectors, a sequence of the respective nearest neighboring vectors for each already existing entry in the set of neighboring vectors is entered in turn in the set of neighboring vectors, continuing until no new entries are received or until entry is terminated because a number of neighboring vectors in the set of neighboring vectors has reached a preset limit.
2. The method of claim 1 , comprising generating a neighboring vector table for a codebook vector in which first a number of the nearest neighboring vectors are identified for at least some of the codebook vectors in the at least one codebook, and modifying the neighboring vector table created as a result by repeating the following steps:
determining two nearest neighboring vectors to a codebook vector in the neighboring vector table and designating them as first and second parent vectors for that codebook vector;
determining two neighboring vectors next nearest to the first parent vector in the neighboring vector table and designating them as offspring vectors of the first parent vector;
deleting the second parent vector from the neighboring vector table and replacing it with its nearest neighboring vector in the neighboring vector table, provided that one of the two offspring vectors from the first parent vector matches the second parent vector;
determining two neighboring vectors nearest to the second parent vector in the neighboring vector table and designating them as offspring vectors of the second parent vector; and
deleting the first parent vector from the neighboring vector table and replacing it with its nearest neighboring vector in the neighboring vector table, provided that one of the two offspring vectors from the second parent vector matches the first parent vector.
3. Reducing the magnitude of a coding error in at least one voice coding and video coding, comprising vector quantizing a feature vector according to a method for vector quantization of a feature vector, comprising:
searching at least one codebook comprising multiple codebook vectors to find a code vector representing a feature vector;
during the search testing a sequence of codebook vectors to determine their suitability to represent the feature vector; and
for at least one of the codebook vectors that can be tested during the search for the code vector representing the feature vector, determining a set of neighboring vectors before the beginning of the search and including in the search for the code vector representing the feature vector at least some of the neighboring vectors; and
wherein the feature vector is indicative of at least one of voice data and video data, and wherein the reduction of the magnitude of the coding error is relative to a magnitude of coding error when a feature vector is not quantized according to the method.
4. The method of claim 1 , wherein said feature vector is at least one of a voice signal and a video signal.
5. The method of claim 2 , wherein the nearest neighboring vectors are determined by measuring their distance from a codebook vector.
6. The method of claim 1 wherein the feature vector is a voice signal.
7. The method of claim 6 further comprising encoding the voice signal based on the code vector that was found to represent the feature vector.
8. The method of claim 7 wherein the encoding of the voice signal is compliant with standard G.711.
9. The method of claim 1 further comprising encoding a voice signal based on the code vector that was found to represent the feature vector.
10. The method of claim 1 wherein all of the neighboring vectors are included in the search for the code vector representing the feature vector.
11. The method of claim 2 wherein the feature vector is a voice signal.
12. The method of claim 11 further comprising encoding the voice signal based on the code vector that was found to represent the feature vector.
13. The method of claim 12 wherein the encoding of the voice signal is compliant with standard G.711.