Research on Sparse Remote Sensing Matching Method Based on LoFTR Descriptor

As an important means of obtaining information about the Earth's surface, remote sensing mapping is widely used in fields such as military, environmental monitoring, resource surveying, and disaster monitoring. Consequently, remote sensing image matching has always been a prominent area of research in image matching. Optical images closely align with human visual perception of the world, but they are susceptible to environmental influences, which can result in information loss or operational failure. LoFTR, as a dense matching model, is limited by its complexity and is not suitable for most equipment conditions. Therefore, this paper proposes a sparse remote sensing matching method based on the LoFTR descriptor, which utilizes RIFT feature points to extract the corresponding LoFTR feature map to obtain the corresponding feature descriptors, and performs registration by calculating cosine similarity. In the experiments, multiple sets of Optical and SAR remote sensing images were used as test data. The experimental results demonstrate that the proposed method effectively retains the advantages of LoFTR, reduces the model's complexity and parameter scale, and achieves good results in remote sensing image registration for both Optical and SAR types, with sub-pixel registration accuracy.