Shape reconstruction from stereo images of fracture surfaces, which are obtained by scanning electron microscopy, is applied to research in micromechanics and fracture mechanics. The shape is reconstructed by matching the corresponding points on a pair of stereo images. To deal with the problem of false-target matching and to reduce computation time, the `interesting points', which usually are pixel locations of important features in each pair of images, are determined first, and the matching is only performed on these `interesting points'. Two algorithms are used for finding the `interesting points'. The first one selects the pixels that have the maximum value of local entropy. The second one selects pixels having the local maximum gray scale value after the image has been processed by a Laplacian of Gaussian filter. In order to reduce the effects of shadows, the logarithm of intensity is calculated before the image is filtered. The algorithm of matching corresponding points in the stereo images is based on the theory of cross correlation. To reduce the false-matching, a coarse and fine matching algorithm is developed. Finally, the height of the surface is calculated by applying stereo triangulation theory.