The registration of serial images is a crucial step in the three-dimension reconstruction of composite materials to obtain the information concerning three-dimension microstructural geometry. Based on the obtainment of grain contours by using wavelet edge detector and the level set method，high curvature points of the plane curve were calculated as the prospective points. An adaptive method was adopted for the polygonal approximation of the digitized raw contours. Instead of setting a fixed length of the support region in advance，the new method computed a suitable length of the support region for each point to find the best approximated curvature. The dominant points were identified as the points with local maximum curvatures.In the second stage，a novel robust point registration algorithm was adopted by maximization of entropy and mutual information. A new energy function，including the joint probability matrix and spatial mapping matrix，was derived. By minimizing the energy function，the estimation of spatial mapping parameters and joint probability matrix were estimated.