In order to improve the registration precision and robustness of image splicing, an improved random sampling consensus algorithm based on local pixel matching was proposed. After completing image feature extracting and feature matching with the scale invariant feature transform operator or other operator, using the local pixels that are independent of feature matching points, then the optimal configuration of the four pairs of feature matching points, and the best homography matrix were determined by matching the local pixel of the reference image with the mapped local pixel of the image to be stitched. The experimental results show that, compared with the classical random sampling consensus algorithm, the computation time is close, the homography matrix is more accurate, and the image mosaic is more robust.