Research On First-order Target Detection Algorithm For Remote Sensing Images Based On Feature Pyramid Contextual Information

The quantity and quality of remote sensing images used to characterize various objects on the surface have been greatly enhanced by the extremely rapid development of remote sensing technology.Target detection tasks are crucial for image interpretation,especially in areas such as intelligent monitoring,urban planning,precision agriculture,and geographic information system(GIS)updates.In recent years,these needs have also driven great efforts to develop various target detection methods in remote sensing optical images.However,unlike ordinary optical images,optical remote sensing images usually present characteristics such as large size differences between targets of the same category,certain similarities between targets of different categories,and complex scenes.The characteristics increase the difficulty of detection,and if the target detection algorithm of ordinary optical images is directly applied to remote sensing images,it is difficult to achieve satisfactory results.Therefore,how to efficiently and accurately detect the targets in optical remote sensing images is an open and challenging problem.Based on the framework of first-order target detection,this paper introduces theories and methods such as feature pyramid network,context-aware module and a ttention mechanisms to address inter-class similarity,intra-class diversity and scene complexity in remote sensing image target detection,taking the problems of how to effectively obtain rich feature information and how to quickly filter out high-quality feature information from a large amount of complex scene information as the starting point.The main research works are as follows:(1)For the problems of inter-class similarity and intra-class diversity of remote sensing images,a remote sensing target detection method based on cross-scale feature fusion pyramid network is proposed.First,a cross-scale fusion module(CSFM)is constructed for cross-scale fusion in order to extract sufficiently comprehensive semantic information from feature maps of different scales.Secondly,a deeper feature multi-level fusion is performed by refining the U-shaped feature pyramid module(TUM).Next,a channel attention mechanism is used to localize to the information of interest and suppress useless information.Finally,the Focal Loss function in the loss function is used to control the large number of negative samples generated during the feature fusion process to achieve the precise localization of the target to be detected.(2)A bidirectional feature pyramid network based on scene context-awareness is proposed to address the problem of complex optical remote sensing image scenes and how to effectively achieve multi-scale feature fusion.The network first uses the designed context-aware module(SCM)to aggregate the contextual information of different regions by pooling at different scales,so as to obtain richer global contextual information.Then,in order to make the feature information at different scales to be effectively weighted and fused,this paper introduces a bidirectional feature pyramid structure to obtain more accurate feature information.Compared with the current mainstream remote sensing image target detection methods,this method achieves better results.