Research On Remote Sensing Image Ground Target Detection Algorithm Based On Deep Learning

Remote sensing image ground target detection can not only help commanders obtain accurate battlefield information in the field of military warfare,but also locate and strike enemy targets and provide real-time target information.It is also widely used in civil fields such as environmental management,regional planning,and mineral resource exploration.Ground targets in remote sensing images include(vehicles,airplanes,buildings,bridges,etc.)and road targets,For the first target,because the internal features of the same target are very different,the detection algorithm lacks good feature extraction capabilities.The target is limited by the shape characteristics of the target,and the detection algorithm cannot completely extract the target contour.Therefore,how to accurately detect the target of interest from the remote sensing image and ensure that the algorithm has a certain promotion ability is the key to the research.In this paper,the following research work is carried out for the detection of ground targets in the above two types of remote sensing images:(1)In-depth analysis of the differences between the two types of targets in remote sensing images and the similarities and differences of corresponding feature extraction,the characteristics and difficulties of corresponding target detection,the advantages and disadvantages of various methods in target detection,and the evaluation indicators in target detection?(2)For aircraft and vehicle targets,a ground object detection method for remote sensing images based on integrated residual network is proposed.For the problems of variable target size,rotation,and complex background,Faster RCNN is used as the basic network,and Kmeans ++ clustering algorithm is used Design the target reference frame size,select a more complex residual network as the basic feature extraction network,add a background classification subnetwork to the network to help distinguish the target and background,build an integrated network model,and the experiment proves the effectiveness of adding each module.The detection accuracy has reached more than 89%.(3)Aiming at road targets,a full-convolution network based remote sensing image ground object detection method based on deformable pooling and cavity convolution is proposed.To address the problem of different road target shapes,the FCN network is used as the basic network to design The deformable pooling kernel extracts road target features,adds batch standardization layers,replaces traditional feature extraction methods with hole convolution,and uses the channel feature analysis module to further improve network performance.The experiment proves that the addition of each module is conducive to the improvement of road target detection accuracy,and the detection accuracy reaches more than 84%.