Research And Application Of Multi-user Collaborative High-resolution Remote Sensing Image Rotating Object Extraction Method

With the rapid development of remote sensing technology,the resolution,data volume,and information content of remote sensing images have greatly improved in recent years.Now,a large number of high-resolution remote sensing images can be provided,which not only provide spatial information but also capture subtle texture changes.Analysis and recognition based on high-resolution remote sensing image technology have now been applied to remote sensing image target detection tasks.At present,many applications in the field of remote sensing,especially those related to major engineering projects,rely heavily on manual interpretation.However,there are challenges in the remote sensing field such as massive data volume,complex data quality,and high data processing difficulty,which lead to high costs for manual interpretation.To fully utilize remote sensing data,conducting research on remote sensing intelligent interpretation is a meaningful direction.Although object detection algorithms based on deep learning have been widely used in the natural image field,there are still many problems that need to be solved in the remote sensing field,including: The size of remote sensing images is huge;The direction of the objects is arbitrary;The problem of small objects;High complexity of the background;Lack of training samples,as well as the problem of data silos that exist due to interests and privacy concerns.To address the aforementioned issues with deep learning-based object detection methods in the remote sensing field,this thesis mainly contributes in the following aspects:(1)A high-resolution remote sensing image rotation detection dataset,DRD-GS,was created for the specific class of rotating objects.Firstly,high-resolution remote sensing images that meet the scene requirements were selected,and then pre-processing operations such as orthorectification and image fusion were performed on the selected images.Next,a high-efficiency and fine-grained annotation method was applied to annotate arbitrary quadrilaterals on the cropped panchromatic images.To solve the problem of the large size of remote sensing images,the images were sliced.DRD-GS currently includes two classes,greenhouse and ship,and contains 123,115 annotated instances.Traditional object detection algorithms were used for benchmark testing,verifying the effectiveness of DRD-GS.(2)A dual-mode rotation regression network,DRRN,was designed,and a combined regression loss function was proposed to improve the traditional smooth L1 loss.Firstly,in response to the problem of arbitrary object orientation in remote sensing images,DRRN was designed,which mainly consists of two parts: dual-mode region proposal network and rotation regression network,and can use two different forms of anchors,horizontal or rotational,to achieve two detection modes.DRRN adopts the rotation intersection over union(Io U)calculation method,and uses rotation non-maximum suppression(NMS)as the post-processing operation based on the rotation Io U calculation.According to the characteristics of different class aspect ratios in the dataset,different rotation NMS thresholds were set for greenhouse and ship.At the same time,a combined regression loss with Io U was proposed to replace the traditional smooth L1 loss,so that it can be consistent with the evaluation standards of detection.The effectiveness of the method has been demonstrated through comparative experiments.(3)A bi-directional cross-layer connection feature pyramid network,bc-FPN,was proposed.To address the issue of small objects in remote sensing images,a bi-directional cross-layer connection feature pyramid network,bc-FPN,was proposed to fully propagate bottom-level position information and top-level semantic information.This module has good portability and features plug-and-play functionality.The effectiveness of the module has been demonstrated through ablation experiments.(4)A supervised hybrid attention mechanism,SHAM,was proposed.To address the problem of complex background in remote sensing images,a supervised hybrid attention mechanism of tandem connection,SHAM,was proposed to suppress image noise and enhance obbject information.This module has good portability and features plug-and-play functionality.The effectiveness of the module has been demonstrated through ablation experiments.(5)A strategy of multi-user collaborative training was proposed,and a remote sensing image interpretation system was designed and implemented.To address the problem of insufficient training samples and data silos,a strategy of multi-user collaborative training was proposed,in which multiple participants can jointly train the model without sharing data,and the model’s performance can be improved through collaborative modeling,while ensuring data privacy and legal compliance.Based on the research on the project requirements and the current situation at home and abroad,a remote sensing image interpretation system was designed and implemented,which can perform various functions such as remote sensing image management,remote sensing image querying,intelligent interpretation of remote sensing images,and multi-user collaborative training.The usability of the system has been validated through system testing.