Abstract: A novel algorithm for multi-image matching by using a manifold learning method is presented.The matching algorithm applies the Laplacian Eigenmap algorithm to map feature points from different images to the same embedding space.Meantime,the local and global distribution similarities of feature points are calculated by SIFT descriptor and location information.Finally,a series of carefully designed experiments on two groups of the wide-baseline image sequences are designed to demonstrate and validate the performance of the proposed algorithm,which is higher than that of the LE-SIFT,SVD-SIFT and LLE-SVD methods in multi-image matching and stereo matching under the multi-image constrains.