The work aims to propose an improved SIFT image watermarking algorithm with respect to the mismatching caused by the image watermarking algorithm robust to geometric attacks based on scale invariant feature transform (SIFT) when the images are matching. Firstly, the carrier image was subject to two-stage non-subsampled Contourlet transform (NSCT) and the watermarking information was embedded into the directional subbands with low-frequency coefficient through mean-value quantitative algorithm. When the images are matching, the generalized Hough transform was used to remove the mismatched feature point pairs to improve the matching accuracy. The geometric distortion parameters were estimated according to the screened feature point pairs. The carrier images with watermarks were subject to geometric correction and watermark extraction, so as to achieve the watermarking robust to geometric attacks. The experimental data indicated that, after the matching pairs of the original feature points were screened, the matching precision was improved at least by 6.8% and the average NC values of extracted watermark were all above 0.9. The results show that the improved algorithm proposed herein can not only resist the geometric attacks and normal signal attacks, but also guarantee the invisibility of watermarks.