A survey of deep-learning-based image matching algorithms

In the fields of remote sensing and surveying, image matching technology is at the core of many applications, especially in geographic information systems (GIS), terrain modeling, and change detection. However, traditional image matching methods rely on manually designed features, which are vulnerable to image deformations, lighting changes, and modality differences, leading to poor matching accuracy and robustness. With the rapid development of deep learning, the performance of image matching has been significantly improved, particularly in remote sensing and surveying. Unlike traditional methods, deep learning does not rely on manually designed features or prior knowledge from the designer, and it can automatically learn meaningful image representations. This enables deep learning to surpass traditional algorithms in terms of accuracy and robustness, especially in handling complex deformations and dynamic environments in remote sensing images. This paper first summarizes the deep learning-based image matching algorithms that have stood out in the past decade and categorizes them into three types: feature detection and description-based image matching algorithms, joint feature detection and description-based image matching algorithms, and self-attention mechanism-based image matching algorithms. The paper then provides a detailed discussion of the development history, framework principles, performance evaluation metrics, and summarizes the advantages and limitations of these three categories of algorithms. Following this, the paper reviews the application scenarios of these algorithms in remote sensing and surveying. Finally, the paper looks forward to the development trends of deep learning-based image matching algorithms in remote sensing and surveying and proposes key research directions for the future.