Abstract: The SIFT descriptor has been widely used in the field of computer vision thanks to its
various invariant attributes; however, its high dimensionality results in redundant data and makes it
time-consuming for application. Therefore, a novel algorithm, considering the inner relationship
between gradient vectors in SIFT descriptor, is presented in this paper, which utilizes the principal
component analysis method based on cosine kernel function. First, a principal component matrix,
which is used to compute the principal direction of the projection matrix, is generated by using cosine
kernel function to extract SIFT descriptors from the sample images. Then, the projection matrix is
applied to the dimensionality reduction of the SIFT descriptors from the new images. In the
experiment, we evaluate the performance of descriptors by means of image matching. The results
indicate that our method can efficiently reduce the dimensionality and also obtain more matches
without sacrificing the matching accuracy and meanwhile improve time performance.

Key words: pattern recognition, image registration, local descriptor, principal component analysis;
cosine kernel function