面向目标检测的空间观测图像精确配准

Precise registration of space observation images for target detection

针对空间目标检测对图像配准精度的要求,本文组合Fourier-Mellin变换和加速鲁棒特征变换(SURF)算法来实现对空间观测图像的精确配准,并研究了该配准算法所涉及的变换模型、特征提取、特征匹配和配准精度等问题。该算法首先利用Fourier-Mellin变换计算图像旋转角度的整数值,根据整数角度将待配准图像反向旋转;然后利用SURF算法检测两幅图像的匹配特征点;最后利用最小二乘方法计算旋转角度的浮点值和平移量,而整数角度与浮点角度之和就是待配准图像的实际旋转角度。文中分析了SURF特征点检测与图像尺寸、DoH响应阈值以及尺度空间层数之间的关系。实验结果表明:提出的算法对图像旋转角度估计值的均方误差为0.0077°,50组实拍图像中星点质心均方误差的平均值为0.1353pixel,能够满足空间目标检测对图像配准的精度要求。

For the requirements of space target detection for image registration,an observation image registration algorithm based on Fourier-Mellin trans-form and Speed Up Robust Feature （SURF） transform was proposed.The transform model,feature detection,feature matching and the matching accuracy related to the algorithm were all discussed.Firstly,the Fourier-Mellin transform was used to calculate the rotation angle,the image to be matched was reversely rotated according to the integer angle of the image.Then,the SURF was used to detect the matching feature points between two images.Finally,the floating point rotation angle and offset were calculated with the least square method.The sum of integer angle and floating point angle is an actual rotation angle of the image to be matched Furthermore,the relationships between SURF feature point detection and image size,DoH response threshold and scale space decompose layers were analyzed.The experimental results show that the Root Mean Square Error（RMSE） of rotation angle estimation is 0.0077°,and the average value of RMSE of 50 groups of star centroids is 0.1353 pixel.The algorithm can meet the requirements of space target detection.