In order to compose the wide visual angle and high resolution image from the sequence of images which have overlapping region in the same scene quickly and correctly, an improved SIFT algorithm which based on D2oG interest point detector was proposed. It extracted the image feature points and generated corresponding feature descriptors by improved SIFT algorithm. Then, feature point matching pairs were purified using the random consistency(RANSAC) algorithm and the transformation matrix H was caculated. Last, the seamless mosaic of images was completed by using the image fusion algorithm of slipping into and out. The images which had the four typical transformations were separately processed with the traditional SIFT and the proposed method in the experiment. The result indicated that the number of feature pairs is fewer, the mosaic time is shorter and the matching efficiency is higher than that of SIFT algorithm. This method reduces the complexity of operation and improves real-time of image mosaic simultaneously.