Due to the visibility change and surface enlargement in the new view as viewpoint moves, 3D reprojection from just one reference image inevitably produces holes in the destination image. In tackling the problem, a common solution is to conduct image synthesis from multiple reference images. However, how to extract the information needed in the novel view from these images with plenty of redundance information becomes one of the key techniques in image-based modeling and rendering. In this paper, a new method of image synthesis from multiple (currently two) reference images is proposed. The method combines forward warping and inverse warping to fulfil the image composition task for a novel viewpoint. In the algorithm, vision constraint-epipolar constraint is employed to search the pixel from other reference images in filling the hole, so as to avoid depth comparison and re-mapping of the whole reference image. Besides, the searching process is accelerated by taking advantage of pixel ordering on pairs of epipolar lines and the information already in the destination image. Without increasing the complexity of previous warping method, this algorithm greatly reduces the computation required in image synthesis from multiple reference images.