Research On Application Of Remote Sensing Image Registration In Aircraft Navigation Based On Network In Network

Remote sensing image refers to the image obtained by remote sensing such as aerial or satellite.As an indispensable part of remote sensing image processing,it is the core problem of this research to study the remote sensing image alignment algorithm with good robustness and high real-time performance.Based on deep learning,proposes two methods for multi-view remote sensing image registration.method.(1)In multi-view remote sensing image registration,some methods suffer from low accuracy or low real-time performance in complex scenes.To address these shortcomings,this method proposes an end-to-end image alignment algorithm incorporating a Network in Network as a feature extractor with a dual-attention mechanism.The algorithm is divided into three parts: feature extraction,feature matching,and parameter prediction.Firstly,we use network-in-network to improve the extraction ability of the model for complex features,and introduce a dual-attention mechanism to improve the discrimination and localization of features.The experiments show that the alignment accuracy is improved by more than 10% on average and the speed is improved by at least 20% compared with the traditional algorithm,which improves the effect of multi-view remote sensing image registration.(2)The feature extraction algorithm based on deep learning can better understand the highlevel semantics of the image,while the method based on feature points mainly relies on the underlying pixel information of the image.Introduces a multi-view remote sensing image registration system that can effectively combine the image Low-level pixel information and high-level semantics.It includes two parts,dense and sparse registration network.First,the dense registration network is used to extract the multi-scale features of the two images,the multi-scale features are matched to generate a multi-scale feature matching map,and the neighborhood consensus network is used to eliminate errors and generate dense image registration parameters;on the other hand,through The sparse registration network performs sparse feature extraction on two images,uses graph neural network to perform feature matching,and generates sparse matching parameters;finally,the dense image registration and sparse registration parameters are merged to generate the final registration parameters.Experiments show that in the registration of complex multi-view remote sensing images,compared with the first method,the registration accuracy can be improved by more than 10% on average.