Linear block prediction with source classification for image encoding applications

Discusses the results of linear block prediction of images. An image can be considered to be zero-mean real vector process by removing the mean vector. Vectors of dimension k=r*r are formed from subblocks of size r*r, and a subblock is predicted as a linear combination of p, q previous blocks along the row and column respectively. The coefficient matrices are chosen so as to minimize the mean square error over a given prediction frame of N*N pixels. The authors deal only with prediction along the rows. A Levinson type recursive algorithm can be used to obtain the coefficient matrices. As an application to image coding, the residual vectors are encoded using a vector quantizer (VQ) in a closed-loop fashion forming a vector DPCM (VDPCM). The prediction frames are classified according to their mean and variances using K-means algorithm and codebooks are generated for each individual class. This source classification will utilize the codebook more efficiently and result in lower data rate or better quality.<>