In this paper, a multiple description image coding scheme with classified method is presented, in which the wavelet coefficients in each class can be processed separately according to its own source distribution. All the classes will be divided into two larger classes again. One of them is quantized by MDSQ and the other is sub sampled by quad tree and quantized subsequently. The two parts are combined to form two descriptions. It is the number of class quantized by the above two ways that determines the tradeoff between side and central performance. The whole scheme is applied to natural images and simulation demonstrates that the system outperforms other related works. Most important of all, by employing the classified scheme, the central performance can get improved even when the side performance is close to its corresponding single description one.