This paper proposes an adaptive hierarchical vector quantization method for image coding. First, a basic codebook and the codes are generated adaptively. Then the codes are used to form the second layer codebook. Finally, the codes are coded by the second layer codebook and form a multilayer code sequence. This hierarchical method makes use of the structural features and correlations in larger regions and thus remove much redundance in the codes obtained by a single layer VQ method. A much lower bit rate can beachieved without causing any more distortion and more computation.