Research on a kind of remote sensing registration algorithm based on improved SIFT

Abstract

This paper offered an improved SIFT to deal the problems in Scale Invariant Feature Transform (SIFT) algorithm applied in remote sensing image registration, such as one's of massive data computation in the processing of registration, more feature categories is much with lower duplicate degree, complex feature extraction and matching process longer time-consuming, not matching the requirements of real-time, being sensitive to changes of gray value characteristics. The new SIFT algorithm speeds up the registration, making it more adapt to remote sensing image registration, stronger real-time performances. In this paper, the main issues are as follows: in the first place, in order to get rid of the noise and enhancing the effect of contour features, Fourier transform the pre-process remote sensing image date in remote sensing image, and time the result with high pass filtering. Secondly, establish differential Gaussian pyramid by DOG operator, and compute extreme points. And screen extreme points to reduce unnecessary points to deal with the sensitivity to gray value characteristics changes. We use the average value of 8 adjacent pixels to substitute the original extreme point in responsing to sensitivity of SIFT to the gray value characteristics changes. Using 128d vector to descript the feature points, so calculation by original algorithm is larger and lower. We prefer eight affine form concentric circles instead of 4 ∗ 4 board within the scope of 16 ∗ 16 pixels adjacent to the key points, making description of the key points reduced from the original 128 d to 64 d and improving the efficiency of the key points matching calculation. In order to pursue the best matching key points in the process of image registration, we use the two sets of remote sensing image's feature points get the pair set of optimized matching key points, then put the set in remote sensing image registration. Experiments with Gaofen-2 satellite remote sensing image data showing a higher computing speed and registration efficiency than the original algorithm.