Abstract: To overcome the redundant feature descpriptor, high computational complexity, low matching precision when the angle of rotation or view spends greatly in image registration method based on SURF algorithm, an improved SURF algorithm is proposed. First, the feature points are extracted using Hessian matrix. Then, the feature descriptor for each keypoint in the circular neighborhood is constructed using Haar wavelet response; meanwhile, the normalized gray values difference and second-order gradient of this region are computed. Finally, RANSAC algorithm is applied to eliminate false matches. This method not only performs faster than SURF algorithm, but also fully employs the image gray information and details to acquire higher accuracy. Results indicate that the proposed method has strong robustness and stability for blur, illumination difference, angle rotation and viewpoint change. A well-edge mosaic image is obtained without obvious geometric misalignment in the remote sensing image mosaicking process. This method is an effective image registration algorithm with high-speed and precision, and it satisfies the need of registration in the remote sensing image mosaic.

Key words: computer application, image registration, feature extraction, SURF algorithm, second-order gradient