Aiming at the problems of low matching accuracy and poor real-time performance of traditional local feature matching algorithms in complex scenes, an image matching method based on CenSurE star fusion of marginalization outliers is proposed in this paper. Firstly, fast bootstrap filtering preprocessing is performedon the template image and the image to be matched. Subsequently, an adaptive threshold based on CenSurE star algorithm is proposed for feature detection. Secondly, for the first time, the BEBELID (Boosted efficient binary local image descriptor) descriptor is used in conjunction with the improved CenSurE star algorithm to obtain efficient binary descriptors using machine learning based classification methods. Finally, MAGSAC++ (Marginalizing Sample Consensus) algorithm is introduced to marginalize outliers and obtain spatial geometric transformation relationships, eliminating errors in preliminary matching and improving matching accuracy. Through the experimental comparison of the standard Oxford dataset, compared with the BRISK, ORB, AKAZE, and the traditional CenSurE-star algorithms, this method has a more uniform distribution of feature points, fewer mismatched points, and possesses stronger robustness in terms of blurring, illumination, point-of-view, and scale variations, which improves the matching accuracy of the algorithm in complex scenes and further enhances the real-time performance.