existing multi-scale ORB algorithms having low accuracy and long operation time,an improved algorithm is proposed.Firstly,a multi-scale space is established by down sampling and filtering to extract ORB feature points,the nearest neighbor method is used to match feature points,and the significance of feature points is introduced into cosine similarity to screen matching points by using the similarity of gray distribution in the neighborhood of matching points.Combined with the local entropy of the image,the K-means algorithm is improved,and the initial class center point is detected only in the area with dense feature points,so that the feature points can be classified more reasonably according to the image.Then,the improved PROSAC algorithm is used to calculate the transformation matrix between images,which further reduces the computational complexity while ensuring the accuracy of the transformation matrix.Finally,the performance of the algorithm is verified by two groups of experiments.