Abstract: A graph theory-based mismatching detection method is proposed in this paper. At first, two complete graphs are constructed by the correspondences in left and right images, respectively. Then an induced graph is constructed by using the sum of similarity of the triangles corresponding to each node in complete graphs as the attribute value. Finally, the induced graph is refined by removing the node of which the attribute value is the smallest in the graph. In order to automatically locate multiple mismatching points, the graph theory-based mismatching elimination method is a recursive process. The whole process scheme is as follows, complete graph building, induced graph building, and mismatching point locating. The experimental results demonstrates that the accurate mapping model between matched points is not necessary in our mismatching detection method, while the local simila-rity of triangles is sufficient for locating the mismatching points. In addition, the true positive rate is higher and the false positive rate is lower compared to classical RANSAC (random sample consensus) bundler detection method.