Fabric image registration based on affine transform and Levenberg-Marquardt algorithm

Abstract

针对多光谱成像系统采集的织物图像的颜色色差分析问题,提出了一种基于仿射变换与LevenbergMarquardt(LM)算法的图像配准方法。从配准角度出发,利用提出的配准方法将标样图像与打样图像配准后进行空间色差分析。多光谱系统采集的织物图像的形变,包括平移、旋转、缩放和错切变换,符合典型仿射变换模型。提出一种新的方法来估计仿射变换矩阵,该方法对两幅图像的对数极坐标幅度谱积分曲线进行匹配,将仿射变换矩阵求解转化为一个非线性最小二乘拟合问题,进而利用LM算法搜寻最优参数值,同时引入分块配准以得到更好的配准效果。实验结果表明,与传统基于Fourier-Mellin配准算法和基于尺度不变特征变换的特征点配准算法相比,提出的算法可获得更好的配准效果,可有效解决具有周期性元素的织物图像配准问题,有助于织物图像色差评估。 In order to solve the problem of color difference analysis for textile fabric images captured by a multispectral system, a new method of image registration based on affine transform and Levenberg-Marquardt (LM) algorithm is proposed. In a registration perspective, the standard image and the batch image are first matched by the proposed method and then taken for analysis of color difference. The deformation of the textile image captured by the multispectral system, including translation, rotation, scaling and shearing, conforms to the classic model of affine transform. To estimate the transform matrix, the proposed method first calculates the integral curves of the log-polar magnitude spectra of the two images for mapping. The original problem of solving the transform matrix is then converted to a numerical problem of nonlinear least squares fitting where the LM algorithm is employed for optimal value searching. Besides, the block registration is introduced to achieve more accurate registration. Experimental results show better registration accuracy of the proposed algorithm compared with the traditional Fourier-Mellin algorithm as well as the scale invariant feature transform based feature point matching method. It also solves the registration problem for textile images with periodic patterns effectively, which contributes to the following process of color difference evaluation.