Feature Matching Of Low Altitude Remote Sensing Images Based On Structural Features And Deep Learning

Feature matching can connect the same or similar attributes in two target images and is the link from low-level vision to high-level vision.As the most basic part of computer vision tasks,feature matching plays an important role in many scenarios.Feature matching techniques have been applied to various kinds of computer vision work,including image registration,image fusion,image retrieval,object detection and tracking,and 3 D reconstruction.It can be said that the quality of the feature matching results directly affects the completion effect of the subsequent computer vision tasks.Low-altitude remote sensing images are mainly collected by small UAVs.Due to the limitation of their flight altitude,the obtained image amplitude is small,the amount of image data is large,and the scale between adjacent images is not consistent,so there are large inclination and large rotation Angle between the images,which makes the images to be registered low and the matching effect of the algorithm poor.Moreover,due to the different shooting instruments or sensors,the resulting images generated will change in luminosity,which makes some intensity-based matching methods ineffective.In addition,with the change of imaging perspective,the extension of shooting time and the change of sensor type,the low-altitude remote sensing images will appear occlusion,complex texture and multi-mode,which makes it more difficult to match the features of the image.Under the influence of the above factors,the low-altitude remote sensing images have a high degree of geometric distortion and the situation is complex,which makes the degree of automatic matching low and easy to appear wrong matching.To address the above high feature-matching error rate in low-altitude remote sensing images,we propose a feature-matching algorithm for point-based structural features and deep learning.This method is divided into a three-step strategy:(1)to reconstruct the initial matching set.First,the mainstream SIFT feature matching algorithm is used for coarse registration to build an initial putative matching set,and then a smooth neighborhood constraint is used to screen out the obvious unmatched point pairs and reconstruct the feature point neighborhood relationship.(2)Visualization of the local structure information.The K nearest neighbors of each feature point are connected in turn,and a series of binary images are generated through the regular grid and the Bursonham midpoint algorithm,and the expression of the neighborhood topology of each feature point is visualized to generate VLSG feature descriptors.(3)Siamese attention network discrimination.For the proposed VLSG feature descriptors we formulate a twin discriminant network with two branches composed of the same specific convolutional neural network and reduce the operation time of the entire algorithm by sharing weights.We add spatial scale attention module and spatial structure attention module to each branch of the twin determination network,through which dynamic scale control enables appropriate structural operations to achieve more precise matching.Our entire feature matching algorithm does not require the texture information of the image,but only the local feature point information of the image.We experiment on the proposed algorithm in several low-altitude remote sensing image data sets,and the results show that the algorithm is significantly better than the other seven advanced feature matching algorithms,and the algorithm has also achieved excellent performance in low-altitude remote sensing image registration.In this paper,the feature matching technology of low-altitude remote sensing images is studied.For the above difficulties,we propose a feature matching algorithm based on point structural features and deep learning.Our method falls into a two-step strategy,first using the mainstream SIFT feature matching algorithm for coarse registration,constructing an initial putative matching set,and then mismatch elimination for this set of matching points.In the process of mismatch elimination,we use a smooth neighborhood constraint to screen out the pairs of obvious mismatch,reconstruct the neighborhood relationship of feature points,and then generate a series of binary images through the regular grid and Burenham’s midpoint algorithm to visualize the expression of the neighborhood topological structure of each feature point.For the proposed VLSG feature descriptors,we develop a special discriminant network,which is a twin architecture with two same branches.Each branch consists of a specific convolutional neural network.We input the feature descriptors of the sense image and the reference image into two branches to realize the similarity measurement between the two feature descriptors,and the two branches reduce the operation time of the whole algorithm by sharing weights.In addition,we also add spatial scale attention module and spatial structure attention module in each branch of the twin determination network,through which dynamic scale control and appropriate structural operations can be realized to achieve a more precise matching.The whole feature matching algorithm does not require the texture information of the image,but only the local feature point information of the image.We tested the proposed VLSG-SANet algorithm in several low-altitude remote sensing image data sets,and the results show that the algorithm is significantly better than the other seven advanced feature matching algorithms,and it has also achieved excellent performance in low-altitude remote sensing image registration.