For the Scale-Invariant Feature Transform (SIFT) algorithm, the eigenvectors at the key points (the extreme points of a stable scale space) are computationally complex and usually have high dimensions, while the values of the polynomial deterministic matrix are limited.To solve the problems, an SIFT image registration algorithm based on the block polynomial deterministic matrix is proposed.The SVM-derived sparse representation method is used to reduce the high-dimensional descriptor vector extracted by SIFT to a low-dimensional sparse feature vector, which reduces the dimensions of key-point description vector.The Euclidean distance is used for similarity measurement of the key-point feature description vector.The contrastive analysis with the traditional algorithms shows that the improved algorithm can effectively improve the registration accuracy and enhance the real-time performance.