The high noise and local deformation of multi-modal images reduce the accuracy of scale invariant feature transform (SIFT) image matching. To solve this problem, a new method based on the SIFT framework, which fuses the phase consistency optimization strategy and the gradient direction of principal component analysis (PCA) with the 8 direction of latitude reduction, is proposed in this paper. This method fuses the histogram of the orientated phase congruency (HOPC) method to extract the direction of the image, and adopts PCA to extract the main direction, which effectively solves the problem that the matching accuracy decreases due to the inversion of the direction of the image. Using the image phase instead of gradient intensity, the difficult problem of direction extraction is effectively solved when image edge characteristics are not obvious. Finally, the random sample consistency (RANSAC) algorithm is used to eliminate false match points. Simulation and experiments show that compared with the SIFT algorithm and PCA-SIFT algorithm, the proposed method improved the number of match points and matching accuracy, significantly reduced the mismatching rate. The statistical results show that the number of match points raised in this paper increases by 20.1% and 200% respectively compared with the former two algorithms.

Introduction

Multi-modal image matching has a wide range of applications in the fields of medicine, remote sensing, and navigation [1]–[۳]. Due to the significant difference in image contrast, image brightness and image texture of multi-modal images [4]. Multi-modal image matching has always been a major problem and has not been completely solved [5], [6]. Multi-modal image matching can be classified into two types according to methods: template matching and feature matching [7], [8]. Based on template matching, gray or edge information of the entire template area is matched, which mainly includes gray similarity matching, gradient similarity matching and the mutual information correlation method [9], [10]. Due to the different generation mechanisms of different image matching, methods based on template matching have high computational complexity, low robustness and unstable performance [11]–[۱۳]. Multi-modal image matching based on feature matching uses feature similarity of images to match images, including points, lines and edges [8], [14], [15]. The literature [16] used shape context descriptions of feature points to match infrared and visible images, which improves the matching stability. The literature [17] proposed a matching algorithm based on linear features and virtual intersection points to match airport images. The literature [18] used edge features and improved Hausdorff distance to achieve infrared and visible image matching.