This study proposes a novel feature-guided Gaussian mixture model (FG-GMM) for image matching, which generally requires matching two sets of feature points extracted from the provided images. The problem is formulated as the estimation of a feature-guided mixture of densities: a GMM is fitted to one point set, in which both the centers and local features of the Gaussian densities are constrained to coincide with another point set. The said problem is solved under a unified maximum-likelihood framework, in which an iterative semi-supervised expectation-maximization algorithm initialized by the confident feature correspondence is also implemented. This algorithm is flexible and has a general scope, which can handle both rigid and non-rigid image transformations. The transformation in the non-rigid case is specified in a reproducing kernel Hilbert space, and a sparse approximation is adopted to accomplish rapid implementation. Extensive experiments on different real images show that the proposed approach consistently outperforms other state-of-the-art methods, which validates its robustness.