In this study, a novel semi-supervised kernel Fisher discriminant analysis (KFDA) based on low density separation geometric distance is proposed. The method employs the low density separation geometric distance as the measure of similarity and thus improves the generalization ability of the KFDA through a large number of unlabeled samples. First, the original spatial data are implicitly mapped onto the high-dimensional feature space by kernel function. Then, both the labeled data and the unlabeled data are used to capture the consistence assumption of geometrical structure based on low density separation geometric distance, which are incorporated into the objection function of Fisher discriminant analysis as a regularization term. Finally, the optimal projection matrix is obtained by minimizing the objective function. Experiments on artificial datasets and UCI datasets show that the proposed algorithm has a significantly improvement in classification performance compared with the KFDA and its modified approaches. In addition, comparison results with other methods on face recognition problems demonstrate that the proposed algorithm has higher identification accuracy.