Feature matching of remote-sensing images: using optimal regularized nonlinear diffusion model

Unmanned aerial vehicle remote sensing image matching techniques are crucial for urban planning and environmental monitoring. However, existing feature matching methods have limitations due to the lack of effective extraction of key diverse features, which hinders the improvement of matching accuracy for diverse ground scenes. To address the above limitations, we propose a remote sensing image feature-matching method using optimally regularized nonlinear diffusion model. Specifically, we first use the accelerated-KAZE (AKAZE) algorithm framework to detect and characterize features within a nonlinear scale space. To improve the adaptation of diffusion to the local image structure, we propose the introduction of a novel Charbonnier regularized nonlinear diffusion model for constructing the diffusion coefficients. We then propose to combine Brute force with the K-nearest neighbor algorithm to eliminate invalid matches. Lastly, we modified the sampling approach of the random sample consensus (RANSAC) algorithm. Instead, we achieve iterative dataset purification using the progressive sample consensus algorithm, complemented by the implementation of a termination criterion for the sampling process. Experimental results show that our method achieves an average matching precision of 82.8%, 80.2%, 82.9%, and 91.2% when the test set undergoes scale, rotation angle, noise blurring, and luminance variations, respectively, outperforming the existing feature matching methods.