图像配准过程中,匹配点位置精度是决定图像配准精度的关键。本文针对航空滤光片阵列多光谱图像因各谱段间存在像点位移而使误匹配点剔除比较困难的问题,提出了一种基于匹配点位置差曲面拟合双阈值剔除方法。首先,选取多光谱图像中间波段图像作为基准图像,利用SIFT算法分别提取基准图像和待配准图像的匹配点；其次,在基准图像匹配点处逐点计算两波段图像匹配点的位置差,构建匹配点Delaunay三角网,利用IDW(反距离加权)算法拟合整幅图像位置差曲面；然后,对位置差曲面进行平滑处理,再分别向上向下平移一定容差范围,构成位置差三维阈值空间；最后,利用位置差三维阈值空间筛选出精确匹配点,并完成图像配准工作。理论分析与实验结果表明：该方法可以有效筛选出航空滤光片阵列多光谱图像高精度匹配点,进而有效提高图像配准的精度。

The high-precision registration between aerial filter array multispectral images is an important guarantee for the subsequent image processing and application. In the process of image registration, the position accuracy of matching points is the key to determine the accuracy of image registration. However, the objects of different strips in the same band image are acquired at different moments, the image displacement between single-band images is large, and the difference of geometric errors of matching points between topographic undulating areas and flat areas in the image is more obvious, the false matching points cannot be accurately eliminated by the global matrix. Aiming at the problem that it is difficult to eliminate the mismatched points in the multi-spectral image of aerial filter array because of the image point displacement between each spectral segment, a new method of double threshold elimination based on the matching point position difference surface fitting is proposed in this paper. Firstly, the intermediate band image of the filter array multispectral images is selected as the reference image, and the matching points in the reference image and the image to be registered are respectively extracted by the sub-pixel-level SIFT algorithm. secondly, the position difference of the matching points of the two bands was calculated point-by-point at the matching points of the benchmark image. and the Delaunay triangulation network of matching points in the reference image was constructed. Then, the position difference surface is smoothed, and the position differences between the matching points of the reference image and the corresponding matching points of the image to be registered are calculated point by point, and then a certain tolerance range is shifted up and down respectively to form a three-dimensional position difference threshold space. Finally, the accurate matching points are selected by using the three-dimensional threshold space of position difference, and the image registration is completed. The three-band composite image of the algorithm-registered image in this paper has clear features and well-defined details，which can meet the requirements of subsequent data processing and application. In order to further illustrate the effectiveness of the proposed algorithm, two datasets of filter array multispectral image are registered and verified from both qualitative and quantitative aspects. False colour composite images: the composite image processed by the proposed algorithm has no obvious pseudo-edges and the features are clear, while there are still obvious pseudo-edges in the comparison algorithm; difference image gray-scale histogram: among the experiments of two datasets，the difference image histogram curve of the proposed algorithm has the largest shift to the left, which indicates that the image registered by the proposed algorithm has the smallest difference with the reference image and the best registration effect. Theoretical analysis and experimental results show that the dual-threshold pointing algorithm based on matching point difference fitting curved surface can effectively screen high precision matching points of aerial filter array multispectral image, and then effectively improve the accuracy of the image registration. The surface fitting to the position difference of matching points reflects the trend of image point displacement in each region of the image, and effectively eliminates the false matching points around the correct matching points according to the feature that the image displacement of same region is approximately same.