Research On Object Detection Of Aerial Remote Sensing Images Based On Deep Learning

In recent years,remote sensing imaging technology has been developing rapidly,satellite and aerial remote sensors can provide people with more and more optical remote sensing images with extremely high resolution.These images contain more and more spatial information with the increase of resolution,,which enables people to more intuitively display the color,size,shape,texture and other characteristics of surface targets,andmake it very convenient to observe through the human eye.The application of aerial remote sensing technology in urban planning,land use,traffic guidance,military monitoring and other fields is of great significance.As a basic technical means in the field of computer vision,object detection has been widely explored and deeply studied in natural images.Especially after the development of deep learning technology,the detection performance of object detectors has been greatly improved.However,compared with natural images,remote sensing images have the characteristics of complex backgrounds,large scale changes of target objects,and arbitrary directions and so on,which brings large challenge to the current target detection algorithm for remote sensing image In view of the above characteristics,in order to improve the target detection performance of aerial remote sensing images,the work of this paper mainly includes the following contents:（1）Aiming at the complex background and arbitrary direction of aerial remote sensing images,this paper proposes a target detection algorithm for aerial image rotation frame based on attention mechanism.This paper proposes a novel channel semantic extraction attention mechanism（Channel Semantic Extracting,CSE）,which can extract the relationship between feature channels,redistribute channel self-attention to features,effectively capture global semantic information,and reduce background noise.interference.In the first stage of the algorithm,the anchor-free target detection method is used to avoid complex and redundant anchor frame design,and an angle refinement module is designed to align the angle between the detection frame and the target by calculating the offset.The second stage improves the Fast R-CNN algorithm,adding angle parameters to make the detection results more accurate.After a large number of comparative experiments,the results of mAP of 77.71% and 89.78% are achieved on the DOTA and HRSC2016 datasets,respectively,which is better than other target detection algorithms for aerial remote sensing images,proving the superiority of the proposed algorithm.（2）Aiming at the large scale variation of target objects in aerial remote sensing images,this paper proposes a target detection algorithm based on improved S2A-Net for aerial remote sensing images.Through the analysis of the S2A-Net algorithm,some improvements have been made to address its problems in the feature extraction stage.First,adding multi-scale convolution in the previous stage of feature extraction solves the insufficiency of standard convolution,and pushes the multi-scale feature fusion process to a more fine-grained level.Second,based on the shortcomings of the common feature pyramid network,we reduced the loss of feature map information of the top layer by adopting the soft REGION of interest selection method and generating the most appropriate final region of interest features.Finally,the CSE attention mechanism is added to the detection head to filter out unimportant multiscale features,improving detection performance.The network is trained and tested on the HRSC2016 public dataset,and the final result can reach 90.03% mAP,and the detection speed can reach 20.1FPS,achieving a high balance between accuracy and speed.