Research Of Hyperspectral Image Registration Based On Deep Learning Method

Image registration is a process of matching or overlapping two or more images acquired at different times,different sensors or different conditions.Traditional registration algorithms that rely on artificial design features,such as SIFT,have achieved large success in the field of general image registration.However,due to the characteristics of remote sensing images,such as large range,clutter,and multi spectra,the hyperspectral images have even more hundreds or thousands of spectral bands,which are registered using traditional algorithms,cannot extract discriminative high-level semantic features.When registering,there are often shortcomings such as fewer feature points to be extracted,more wrong matches,or even false to register.In deep learning field,the convolutional neural networks have good feature extraction capabilities.With the deepening of the network layers,robust high-level features can be extracted.Compared with the artificially designed feature registration algorithm,it is more suitable for hyperspectral image registration.Therefore,in this paper,we use a method based on deep learning network model to register hyperspectral images,mainly network design and optimization of the three processes: hyperspectral data reduction,feature extraction and feature matching to improve the accuracy of image registration.The experimental results show that: compared with the traditional artificial feature registration algorithm,the registration algorithm based on the deep learning network model in this paper can achieve the registration of hyperspectral remote sensing images;compared with the existing deep learning registration algorithm,the feature extraction network can extract features with greater robustness in hyperspectral images;the feature matching strategy optimized in this paper can significantly improve the correct matching rate of feature point pairs.Thus,in this paper,hyperspectral image registration based on deep learning network model is achieved from data preprocessing to model design.The main tasks are as follows:(1)Using deep learning-based data dimensionality reduction network-stacked autoencoder(SAE)to reduce the dimensionality of hyperspectral images,the purpose is to extract feature bands with rich spatial information and input the registration network to reduce network parameters while improve the registration accuracy and efficiency of the network.The experimental results show that,compared with the traditional data dimensionality reduction method PCA and AE,the SAE method in this paper can simultaneously compress the data and obtain dimensionality reduction data with rich spatial and spectral information.(2)An improved feature extraction network is designed.For the characteristics of low spectral resolution and insignificant features of hyperspectral images,the VGG network is used as the basic network to appropriately deepen the number of network layers and add a deep residual network block(Res Net).It makes the network layer deeper and effectively overcomes the problem of gradient disappearance,which can extract highlevel semantic features that are more conducive to image registration and increase its robustness.(3)Optimize the feature matching network,use the joint optimization loss function to jointly optimize the feature extraction and feature matching results,so that the feature matching network results in negative feedback on the feature extraction,guide the feature extraction process,and thus improve the registration of the entire registration network capabilities,while achieving end-to-end image registration without human intervention.