This paper analyses the basic theory of the EZW algorithm, and proposes an improved EZW based on eight symbols instead of four symbols used in EZW. In this method, giving the definition of zero-tree structure anew does not only reduce the complexity of encoding and decoding, but also saves the storage space and improves compression ratio. In e-classifying the significant symbols, the edge information is considered. At the same time, the number of zero tree roots is geatly redued and compression ratio is improved. Experimentsal results show that the improved algorithm performs better than EZW in reconstructing image quality .