In the compression of some particular kinds of image sources

such as human face

vector quantization should naturally be considered to exploit its statistical properties. In this paper

a new vector quantization method in the wavelet transform domain is proposed. We use tree structure to organize coefficients. In each tree

nodes are pruned or retained according to their importance. The survived nodes are serialized to be a vector

which will be the input of vector quantization; a map indicating the positions of these nodes is also stored

which is to be used in decoding. It embeds together the principles of EZW and SPIHT

exploiting fully both the parent children dependencies and brother dependencies. In our algorithm's framework

SFQ's tree pruning algorithm can also be embedded to increase the PSNR of reconstructed images

though we simply choose the threshold pruning method to reduce complexity of algorithm. Using the proposed algorithm

the reconstructed facial image in very low bit rate(about 0 08bpp) is superior to that of EZW

SPIHT

SFQ in both perception and PSNR. The algorithm is very suitable for the compression of particular kinds of image sources.