基于改进的自引导网络的图像配准

Modified Self-Guided Network Based Image Registration

摘要: 图像配准问题关注的是如何通过空间的几何变换将两幅或多幅图像中相似的部分进行对齐。U型卷积神经网络在图像配准任务中有着十分成功的应用。本文首先分析了U型卷积神经网络自身结构的正则性对图像配准的影响，然后以此为出发点改进了一个图像去噪领域的卷积神经网络，将其应用在图像配准任务中。最后在手写体数字数据集的类内图像上进行了配准实验，结果证明本文所提出的方法在可视化效果和数字指标上均有所提升。本文还展示了两种网络的泛化能力，以及多种正则项对于配准数值结果和可视化效果的影响。

Abstract: Image registration is concerned with aligning similar parts of two or more images through spatial geometric transformations. U-shaped convolutional neural networks have been used successfully in image registration tasks. In this paper, we first analyze the effect of the regularity of the U-shaped convolutional neural network structure on image registration, and then improve a convolutional neural network in the field of image denoising as a starting point to apply it to the image registration task. Finally, the registration experiments are conducted on intra-class images of a handwritten digital dataset, and the results demonstrate that the proposed method improves both visualization and numerical metrics. This paper also demonstrates the generalization ability of both networks and the effect of multiple regular terms on the numerical results and visualization of the registration.