Vector quantization

I. How is VQ designed?

An image is divided into different blocks (e.g. 4x4 pixels). We use a codebook of length 16, 32, 64, 256. Sending the codebook, the index of the codebook to the receiver for decompression. Initialization for the codebook consists of 2 different strategies

1. Choose one vector as the averaged vector of the training set. Then split the training set into two sets, four sets and eight sets according to the requirements of the length of the codebook (i.e. 16, 32, 64)
2. Combine two vectors into one if they contribute a smallest distortion, continue until we have the required codebook.

II. What is VQ used for?

-Speech coding
-Image compression in low bit rates