The traditional SURF algorithm using a fixed threshold in image matching has problems of uneven feature points, low matching accuracy and high time complexity. To address these issues, an improved fast image matching algorithm based on the SURF algorithm is proposed. Firstly, through the statistical analysis of the response of the Hessian matrix, an adaptive threshold method is proposed to extract more effective feature points in the image pyramid. Then, the method of quadtree is introduced to homogenize the proposed feature points to reduce the false matching rate. To prevent the quadtree from being over-split, this article proposes an adaptive split depth method to improve the quadtree. Finally, this article combines the BEBLID binary descriptor with the improved SURF algorithm for the first time, and uses the sampling mode based on machine learning to build strong descriptive binary descriptors for feature points, which improves the matching accuracy and enhances the matching speed. Experimental results show that the matching accuracy of the proposed algorithm in the Mikolajcyzk image dataset test is 9. 7% to 27. 0% higher than that of the traditional SURF algorithm, and the speed of algorithm is more than 50% . Compared with SIFT, SURF, BRISK and ORB algorithms, the improved algorithm proposed in this article has better robustness and real-time performance.