Fine Recognition Of Geospatial Targets Based On High-resolution Imaging

In recent years,due to the rapid development of satellite imaging technology,the spatial resolution of remote sensing images has reached sub-meter accuracy and covers a wide area,which can provide rich data information for scene tasks such as military defense and intelligent transportation.For this reason,many rotation object detection algorithms have been developed,which are computationally complex and slow,and rarely involve fine-grained objects.Therefore,this paper studies the fine recognition of geospatial targets based on high-resolution imaging in order to intelligently and efficiently interpret remote sensing images.In view of the inconsistency between classification and regression caused by the angle regression method of the rotation object detection algorithm,the one-stage horizontal bounding box detection algorithm YOLOv5 based on the anchor-based method is selected as the baseline,combined with the circular smooth label method to achieve precise positioning of remote sensing image geospatial targets.The multi-head self-attention is used to improve the learning ability of the network for fine-grained target features,and further improve the accuracy of fine-grained classification.The fine recognition network optimized by the activation function Acon improves the detection accuracy.Experimental results show that the designed YOLOv5＿CSL fine-grained recognition network can effectively alleviate regression and classification inconsistencies,and achieve 38.5%m AP on the high-resolution remote sensing image fine-grained recognition dataset FAIR1M.Aiming at the problem that the rotation object detection algorithm has poor detection results of narrow and long objects and tail objects due to the wide range of target scale changes and the unbalanced number of categories.Select the one-stage rotation detection algorithm S~2ANet based on the anchor-based method as the baseline,and design an adaptive Io U threshold sample allocation module based on the target shape to improve the detection effect of long and narrow targets.Design an improved balanced classification loss function module based on Focal Loss to improve the detection accuracy of tail categories.The experimental results show that the rotation bounding box fine recognition network designed using these two modules can effectively alleviate the above problems,and obtain a m AP of 47.86%on the FAIR1M dataset,which is 1.5%more accurate than the original algorithm.Aiming at the large amount of calculation and small target detection problems in the anchor-based method,the rotation object detection algorithm Oriented Rep Points based on the anchor-free method is selected for implementation to explore a better fine recognition paradigm and improve the detection effect of small objects.Fully compare the performance of the fine recognition network implemented by the anchor-based and anchor-free methods in FAIR1M,select the most appropriate model deployment,and realize the engineering application value of the research.The experimental results show that the rotation bounding box fine recognition algorithm based on the anchor-free method can effectively improve the model’s ability to detect small objects,and the positioning accuracy is higher.