The work aims to propose an improved SIFT descriptor to improve the robustness and distinguishability of SIFT descriptor which produces a large number of mismatches in an image with affine transformation. Hessian Affine detector was adopted to extract feature point. In the characterization phase, the elliptical was normalized into neighboring region to enhance the affine invariance. Scale invariance was improved by affine scale space. The gradient histogram in the circular region was calculated by the log-polar transform to improve the rotation invariance of descriptor. The image with scaling and rotation, change of camera angle, image blurring, JPEG image compression and illumination change was selected, with overlapping error εo less than 0.5. Its recall & 1-precision curve graph display performance increased by 2%~10%. Compared with SIFT, GLOH and PCA-SIFT descriptor, etc., the proposed improved descriptor is more robust and more distinguishable.