Disparity compensation prediction and transform coding are incorporated into a hybrid coding to reduce the bit-rate of multi-view images. However, aliasing and inaccurate displacement impair the performance of disparity compensation, especially in wavelet domain. In this paper, we propose a level-bylevel adaptive disparity compensated prediction scheme for scalable stereo image coding. To get spatial scalable feature, wavelet transform is first applied to the target image of a stereo image pair. A separable 2-D filter applied to the reference image is optimized for each resolution layer by minimizing the energy of the prediction high-bands of the target image. To form a multi-resolution representation, similar processes are then applied to the low-band image pairs generated by the prior resolution layer iteratively. Experimental results show that the proposed scheme can provide significant coding gain compared to other scalable coding scheme.