Deep Learning Based SUAV Low-altitude Remote-sensing Image Registration

Small unmanned aerial vehicles(UAVs)represent a trend in the development of airborne remote-sensing platforms in recent years.Due to their low cost,flexible flight,and high relevance,small UAVs can provide a customizable airborne platform that can be installed with a variety of sensors to quickly acquire high-resolution images in small areas where flying is difficult.When the UAV is monitoring the ground,it is cannot avoid the influence of flight attitude(yaw,pitch,roll)due to the following factors: 1)natural factors such as wind speed and direction,complex terrain;2)human factors such as improper operations,differences in flying height,and speed;and 3)equipment factors such as battery issues and GPS positioning error,which cause the acquired images to be squeezed,twisted,stretched,and offset relative to the target position of the ground.These changes have a strong effect on the accuracy of image registration.To solve the aforementioned multi-view issues and provide a reliable registration for subsequent applications,we propose a novel mismatch removal approach using the Siamese network.The main contributions are listed as follows.1)To solve the problem of unstable inlier matching,lack of generalization ability and difficulty in distinguishing a similar background,a neighbor-guided patch representation is designed for feature point descriptor;2)To solve the problem of weights,model adaptability,nonadaptive optimization parameters,and local image information,this study proposes a Siamese network to measure similarity between the aforementioned patches for mismatch removal;3)To solve the aforementioned multi-view issues,we designed a rotation-invariant layer in the abovementioned learning framework.Secondly,in order to further improve the efficiency of low altitude remote sensing image registration,based on the generative adversarial network of low altitude remote sensing image registration method,its purpose is to achieve an end-to-end mismatch removal algorithm,improve the whole efficiency of the algorithm.Firstly,neighborhood feature descriptors are constructed using generative adversarial networks.In the training stage,two discriminators are designed to strengthen the relationship between the image patch representations of the corresponding feature points in the reference image and the sensed image.Then,the multi-layer fusion network is used to solve the problem that the outliers with similar background could not be mistaken for inliers,and more outliers are eliminated,which improved the accuracy of the algorithm.According to the data set collected by DJI UAV,this paper carries out comprehensive experiments on the above two algorithms.In the experiment,we designed feature matching and image registration experiments for both algorithms,which are compared with the existing state of the art 7 methods.In the second algorithm,we also added a comparative experiment of satellite remote sensing image data sets.The results show that the framework algorithm in this paper gives the best results in most cases.