Research On Image Matching Algorithm Via Spatial Transformation Network

Remote sensing image matching refers to the calibration of one image and another image in the same area so as to match the same pixel in the two images.Image matching has important applications in military field,forest change detection,urban planning and so on.The goal of remote sensing image matching is to develop a remote sensing image matching algorithm with high accuracy,good robustness and high matching efficiency.However,the matching robustness of most existing algorithms is not high enough,and the matching accuracy needs to be further improved.Faced with the problem of poor matching effect of complex images,this thesis proposes two algorithms based on deep learning method to solve the above problems.Remote sensing image matching refers to the calibration of one image and another image in the same area so as to match the same pixel in the two images.Image matching has important applications in military field,forest change detection,urban planning and so on.As an essential part of remote sensing image processing,the research on remote sensing image matching algorithm with high registration accuracy,good robustness and high real-time performance is the core issue of this research.However,the matching robustness of most existing algorithms is not high enough,and the matching accuracy needs to be further improved.Faced with the problem of poor matching effect of complex images,this thesis proposes two algorithms based on deep learning method to solve the above problems.1.Aiming at the problems that some remote sensing image matching algorithms have insufficient feature extraction ability and many mismatched points,a remote sensing image matching algorithm based on cyclic parameter synthesis spatial transformation network is proposed.In the feature extraction stage,an improved spatial attention mechanism is added to focus on the important areas and significant features of the image to enhance the feature expression ability of the network.In the matching stage,a double filtering method of coarse filtering and fine filtering is proposed to improve the robustness and registration accuracy of the model.At the same time,the loss function is improved to improve the generalization and fitting ability of the model.In the experimental results,compared with a variety of advanced methods,the proposed algorithm can obtain higher registration accuracy and lower registration error,and can also obtain better registration results for different types of images.2.Aiming at the problems of low accuracy and mismatching in remote sensing image matching algorithm,a remote sensing image matching algorithm based on feature fusion and enhanced matching is proposed.In this thesis,Involution modified and upgraded ResNext was proposed in the feature extraction stage,and SPANet was fused with Res Next to improve the feature extraction capability of the network.In feature matching stage,a method of enhancing matching is proposed,which uses cross-correlation and nearest neighbor to next neighbor ratio method to filter mismatched points to deal with complex image and background interference.Compared with traditional methods and deep learning methods,this model has achieved good experimental results in many indexes.The two algorithms proposed in this paper are compared with the traditional algorithms and the latest algorithms in recent years on the public data set,and it is proved that the algorithm in this paper achieves superior performance in many indexes.