Abstract:SURF is a scale and in-plane rotation invariant detector and descriptor with better performance, but their stabilities and time complexity are not good enough and unstable features are often detected, which results in needless calculation. The method which extends the detector with information theory and divides the features into sub-collection is proposed to improve performance and matching speed of the algorithm. Firstly detects the maximum point of Hessian around,secondly calculates its information by SURF, then divides the features extracted from both the test and the model object image into several sub-collection, finally the mapping relationship between images is acquired using RANSAC and least squares techniques. The experimental results show that the improved algorithm has the same registration performance but faster speed than SURF.