Research On Object Detection And Instance Segmentation Algorithms In Remote Sensing Images

With the rapid development of remote sensing technology,the spatial resolution of remote sensing images has been dramatically improved,which makes the shape,texture,and other information of objects more clearly displayed.The accurate interpretation of remote sensing images is significant for many applications such as mapping,urban management,and traffic monitoring.However,due to the complex background information and dense instance distribution in high-resolution remote sensing images,manual interpretation of massive data requires a lot of time and effort.Therefore,developing efficient and intelligent remote sensing image interpretation technology is of great significance,which has also attracted widespread attention.As important branches of remote sensing image interpretation,object detection,and instance segmentation in remote sensing images are also the research topics of this dissertation.Although with the development of deep learning,the research on object detection and instance segmentation has made great progress,there are still some important issues to be solved,including the effective selection of training samples,the quality evaluation of training samples,the high-precision prediction of boundaries,and the efficient and accurate prediction of complex shapes.In this dissertation,we will investigate the above four issues.The main contributions of this dissertation are summarized as follows:1.Research on effective selection of training samples in the object detection task.This dissertation proposes a level-wise positive sample selection algorithm(LPS2A)for object detection in remote sensing images.LPS2A first estimates the number of positive samples on each feature pyramid level according to the statistical characteristics of training samples and then selects training samples level by level.In this manner,LPS2A could avoid random matching between objects of various sizes and feature pyramid levels caused by inaccurate model prediction,thus improving the detection performance of LPS2A.Experiments on two public datasets have demonstrated the effectiveness of the proposed LPS2A.2.Research on quality evaluation of training samples in the object detection task.This dissertation proposes a dynamic weighting label assignment(DWLA)algorithm for object detection in remote sensing images.DWLA dynamically adjusts the weights of individual quality scores according to the current model state,so that the quality evaluation indicators could dynamically adapt to the changing model state,thus improving the detection performance of DWLA.In addition,to achieve more stable training,an object-adaptive scheme for constructing the initial candidate of positive samples is proposed.Experimental results on two public datasets have demonstrated the effectiveness of the proposed DWLA.3.Research on high-precision prediction of boundaries in the instance segmentation task.This dissertation proposes an adaptive polygon generation algorithm(APGA)for instance segmentation in remote sensing images.For each instance,APGA generates the corresponding polygon contour by predicting the position of each vertex and the arrangement between these vertices.Based on this manner,APGA could encode the overall structure of the instance and predict more accurate instance boundaries.In addition,a vertex regressor that takes local patches around the vertices as input is proposed to further improve the accuracy of the generated polygon contours.Experiments on multiple datasets have demonstrated the effectiveness of the proposed APGA.4.Research on efficient and accurate prediction of complex shapes in the instance segmentation task.This dissertation proposes an iterative polygon deformation algorithm(IPDA)for instance segmentation in remote sensing images.IPDA evolves the polygon contour from an initial contour based on the geometry of the instance.Compared with the previous methods,IPDA avoids pixel-wise predictions,which greatly reduces the number of model parameters and the inference time.Moreover,to improve the adaptability of IPDA to instances with complex shapes,a novel strategy that could iteratively recover the missing vertices is proposed.Experiments on five datasets have demonstrated the effectiveness and efficiency of the proposed IPDA.In general,this dissertation has conducted four research works on effective selection and quality evaluation of training samples in the object detection task,as well as high-precision prediction of boundaries and efficient and accurate prediction of complex shapes in the instance segmentation task.The experimental results have verified the effectiveness of the proposed algorithms.