Gibbs phenomenon, which occurs in the wavelet-based image compression algorithms under low bit rates, remains an open question for many years. The main cause is that purely-pixel-value-based MSE(mean square error) criteria can not allocate enough bits to the wavelet coefficients corresponding to edges in image. With detail analysis of zero-tree wavelet image compression algorithm originally proposed by Shapiro and then well-modified by Said and Pearlman, the algorithm is improved by suppressing high frequency noises as well as adaptively quantizing coefficients around edges. Experimental results are comparatively given. The main contribution of this paper is the idea of combination of recognition and compression. With the aid of the spatial localization property of wavelet transform, a very flexible bit allocation scheme can be realized, and therefore Gibbs phenomenon is reduced to some extent.