Research On Improving The Geometric Positioning Accuracy Of High-Resolution Remote Sensing Images Based On Intelligent Interpretation

In recent years,with the vigorous development of the global remote sensing satellite industry,mankind’s ability to observe the Earth has reached an unprecedented level.The continuous breakthroughs in satellite hardware technology have made the spatial resolution,temporal resolution and spectral resolution of remote sensing data increasingly improved.Remote sensing images have played an irreplaceable role in social and economic construction.The geometric positioning accuracy of remote sensing images is an important indicator of its quality evaluation,and the registration technology is an important method to improve the geometric positioning accuracy.However,unlike natural images,high-resolution remote sensing image data have the characteristics of multi-source heterogeneous,extremely wide coverage and complex content,which leads to the failure or inaccuracy of traditional feature-based registration methods.In this paper,a number of registration frameworks for high-resolution remote sensing images are proposed,which adopt intelligent interpretation to reduce image complexity and provide prior knowledge for registration tasks,thereby improving the applicability,accuracy and efficiency of the registration method,while achieving the geometric positioning accuracy improvement of high-resolution remote sensing images.Deep learning and convolutional neural network technologies have achieved the most cutting-edge results in many research fields of computer vision due to their powerful ability to learn the essential characteristics of data.Multiple convolutional neural network models are proposed in this paper for intelligent interpretation of remote sensing images,the independent application of which to the remote sensing big data mining and the transformation of remote sensing satellite service capabilities is also of much significance and value.The specific research content of this paper is as follows:Aiming at the registration input mismatch problem caused by the poor positioning accuracy of sensed images,an automatic registration framework based on object detection is proposed which is suitable for high-resolution remote sensing images of urban scenes.Since the rotating object detection can provide more accurate object position information,a single-stage anchor-free rotating object detection algorithm is first designed for intelligent interpretation of remote sensing images,and the main innovations include: First,a novel rotating object representation is proposed by a circle cut horizontal rectangle,which can ensure that the regression parameters do not exceed the definition domain,and avoid the ordering of the vertices,thereby solving the boundary problem and order problem of the current mainstream rotating object representation,while improving the robustness to prediction errors.Second,a lightweight head is designed based on the representation to add the rotating regression to classic benchmarks in an almost cost-free manner.Subsequently,an object region matching process that does not depend on geospatial information is designed,including the selection of object types,object matching strategies based on the rotating object detection results,and the generation of input image block pairs for registration.Finally,the feature-based registration method is used to obtain global homonymy point pairs and calculate the affine matrix.The proposed rotating object detection algorithm has achieved better accuracy and speed than state-of-the-art methods in multiple aerial image and scene text detection data sets;the proposed registration method has obtained comparable accuracy with state-of-the-art methods on test data,while the computational overhead is greatly reduced and the efficiency is improved.Aiming at the problem that the feature-based registration method is not applicable in some remote sensing scenes,an automatic registration framework based on semantic segmentation is proposed in this paper for high-resolution remote sensing image.First,a multi-scale residual fusion and skip connection semantic segmentation algorithm is designed for intelligent interpretation of remote sensing images.The algorithm improves the feature extraction and restoration through the multi-scale residual fusion encoding and decoding blocks.Furthermore,it realizes the full utilization of the information flow,reduces the semantic gap between the encoder and the decoder,sets different degrees of distinction for the feature maps of different stages,and greatly reduces the feature map channels thus improving efficiency through multi-scale skip connection.In addition,it adopts a point rendering iterative up-sampling strategy to effectively improves the boundary fineness of the semantic segmentation result.Subsequently,a corresponding registration process is designed based on the results of semantic segmentation,which effectively circumvents the limitations of feature description operators and reduces the workload of registration.The proposed semantic segmentation algorithm has been compared with state-of-the-art methods on multiple data sets,and has achieved the best in many accuracy evaluation indicators while being rather efficient.The proposed registration method is compared with state-of-the-art methods on experiments,which achieves the best or the second-best accuracy and has an absolute advantage on efficiency.Aiming at the problems of poor positioning accuracy of sensed images and certain features will affect the registration accuracy,as well as the limitations of the registration framework based on object detection and semantic segmentation,an automatic registration framework based on instance segmentation is proposed for high-resolution remote sensing images.First,a single-stage fine-grained instance segmentation algorithm is designed for intelligent interpretation of remote sensing images.The algorithm combines high-resolution feature representation,attention mechanism of bidirectional weighted feature fusion,single-stage instance segmentation strategy,and the post-segmentation processing of boundary refinement into an end-to-end model,aiming to obtain the result of instance segmentation with fine boundary.Then,an instance matching strategy based on the invariant moments is designed to obtain key point pairs,and a cross-validation strategy is designed to obtain global homonymy point pairs,and the registration is finally realized.The proposed instance segmentation algorithm has been compared with state-of-the-art methods on multiple data sets,and it has achieved the best in a number of accuracy evaluation indicators,especially in the accuracy of the segmentation boundary.The proposed registration method has achieved the best accuracy in the comparison experiments with state-of-the-art methods,and its efficiency has an overwhelming advantage.