The work aims to improve the matching accuracy and robustness of the current image matching algorithms. Forstner operator was introduced to accurately extract the feature points of the image. Then, the first-order Haar wavelet was adopted to generate the principal direction of the feature points, and the feature descriptors were generated by obtaining Haar wavelet response. The cosine constraint model was established by means of feature vectors, and the double matching constraint was constructed by combining the distance measurement method of feature vectors, thus completing the matching between feature points. The voting mechanism was introduced to improve RANSAC method. The multiple screening methods were made to remove the false matching points and complete image matching. Compared to the current image matching methods, the proposed algorithm had stronger robustness and matching accuracy. When the total number of features was 200, the correct matching number obtained based on the proposed algorithm was up to 196. In conclusion, the proposed matching technology has better matching accuracy, and has a good application value in the field of identification of package printing products and information security detection.