Abstract: For the poor robustness and real-time problem of large area image registration, a new fast image registration algorithm based on the sub-image features was proposed. First, the sub-image with strong contrast, clear structure was extracted. And then sequentially, the improved Harris detector was taken to extract the corners, the 4-order gradient vectors was used to describe the corners, the improved Euclidean distance was taken to measure eigenvectors' similarity, and the transformation parameters were estimated with Least Squares. Lastly, bilinear interpolation was used to reconstruct the registered image. Experimental results show that the registration accuracy and registration speed of the proposed method are higher than the whole image method, and this idea can be extended to other registration algorithm such as FMT and MI registration methods.