An improved algorithm of KPCA-SIFT for image registration

In this paper, an improved SIFT algorithm with Kernel Principal Component Analysis (KPCA-SIFT) is presented for image registration. Gaussian kernel function is applied to the PCA to extract the principal component for reduced-dimension processing of SIFT descriptor to each feature point. The matched keypoints are selected through similar measure, and the Euclidean distance is replaced by linear combination of cityblock and chessboard distances. The experiments show that this algorithm is robust to image changes in scale, noise and rotation with higher matching accuracy.