Partitioned vector quantizer (PVQ) (Gersho and Gray, 1992) is one of t
he most popular structurally constrained vector quantizers for accurat
e data compression. The design of PVQ involves the partition of the or
iginal vector into a number of subvectors, and assignment of bits to e
ach subcoder for the coding of subvectors. Conventionally, all the pos
sible combinations of partitions and bit allocation patterns, with the
constraint that the total number of bits is fixed, are evaluated in t
erms of the resulting coder performance in order to optimise the perfo
rmance of PVQ. Obviously, this approach is computationally very intens
ive due to subcodebook generations associated with the surprisingly la
rge sum total of all the possible combinations. To solve this problem,
a dynamic resolution analysis (DRA) scheme is presented in this paper
. The DRA scheme employs dynamic resolution (DR), which is a proposed
measure associated with the coordinate variances of the training set a
nd the codebook size, to give an indication of the coding accuracy of
each subcoder, and subsequently predict the overall PVQ performance. T
he optimal or near-optimal design of PVQ is thus obtained without the
need to perform a thorough computer search procedure. The significance
of DRA lis that optimising the PVQ with very large total numbers of b
its and partitions becomes a practical possibility. (C) 1997 Elsevier
Science B.V.