Given that random measurement matrix has defect in hardware realization, a scale-invariant feature transform (SIFT) based on polynomial deterministic matrix algorithm is proposed combining with the sparse projection of compressive sensing theory. The effectiveness of feature vector is enhanced by increasing the numbers of orientation gradient. The dimension of SIFT feature vector is decreased by a polynomial deterministic matrix with the measurement numbers of 7. Accordingly, the Euclidean distance is introduced to compute the similarity and dissimilarity between feature vectors used for image registration, and kd data structure is used to avoid exhaustion. Experimental results show that the proposed algorithm has better performance than the traditional SIFT algorithm and some current modified SIFT algorithms. At the same time, the deterministic matrix is beneficial to hardware implementation of image registration system.