基于VoxelMorph的可见光和红外遥感影像配准

由于可见光和红外的成像机理、成像波段不同，获取的遥感影像之间存在复杂的非线性辐射畸变，传统的配准方法难以实现两者的高精度配准。本文提出一种基于VoxelMorph的可见光和红外遥感影像配准方法，利用卷积神经网络对可见光和红外异源图像进行分步的精细化形变场计算，从而实现快速高精度配准。将可见光图像作为参考图像，利用U-Net网络计算待配准红外图像和参考（可见光）图像的形变场，实现全局对齐的仿射变换，然后通过空间转换网络进一步实现更高自由度变形。采用WHU-OPT-SAR数据集的实验结果表明，与基于尺度不变特征变换（SIFT）算法的传统配准方法相比，本文提出的基于VoxelMorph配准方法可以获得更好的配准效果，验证了基于VoxelMorph的配准方法在多源遥感影像领域的有效性。

Visible and Infrared Remote Sensing Image Registration Based on VoxelMorph

Due to the different imaging mechanism and imaging band of the visible light and the infrared light, complex nonlinear radiation distortions of the acquired remote sensing images can emerge, and it is difficult for traditional registration methods to achieve high-precise registration performance. A visible light and infrared remote sensing image registration method is proposed, based on VoxelMorph, which uses a convolutional neural network to calculate the fine deformation field of visible light and infrared heterogeneous images step by step, thereby achieving fast and high-precision registration. By using visible light images as the reference image, the research uses the U-Net to calculate the deformation field of the infrared image to be registered, and the reference(visible light) image; by realizing global alignment affine transformation, the research further realizes higher degree of freedom deformation through spatial transformation network.The experimental results of WHU-OPT-SAR dataset show that, compared with the traditional registration method based on scale-invariant feature transform(SIFT) algorithm, the VoxelMorph-based registration method proposed can achieve better registration effect, which verifies the effectiveness of VoxelMorph-based registration method in multi-source remote sensing image field.