超融合残差行进几何感知的遥感目标检测

打开文本图片集

中图分类号：TP391 文献标识码：Adoi：10.37188/OPE.20253308.1289

CSTR：32169.14.OPE.20253308.1289

Abstract： This paper proposed an ultra-fusion residual marching geometric perception algorithm，which aimed to solve the challnges of multi-scale，dense overlap，and uneven data distribution in remote sensing image object detection. The hyper-fusion residual marching module optimized the network structure，and its multi-level convolution operation used different scale receptive fields to capture the details of each scale of the object，enhance the model's perception of the object features，and achieve small-scale object feature extraction and large-scaleobjectaccurate positioning.The detection effect was accurately evaluated by calculating the geometric similarity between the detection and the real results，and the fit was carefully considered in the dense overlapping scene of the object，so as to screen the final results，reduce missed detection and false detection，and improve the mAPof the algorithm. A multi-path feature fusion module was designed to fuse different levels offeature information，extract richer object features，enhance network representation and discrimination capabilities，and improve detection mAP and stability. The experimental results on the NWPU-VHR-1O data set showed that mPrecision，mRecall， mAP and mF1 Score were increased by 0.0419，0.1040，0.0455 and O.O85 O，respectively.The experimental results on the RSOD data set show that mPrecision，mRecall，mAP，and mF1 Score are increased by 0.0221，0.1034 0.061 9，and O.O87 5，respectively. The effectiveness and superiority of the proposed ultra-fusion residual marching geometric perception algorithm in the field of remote sensing image object detection are fully proved.

Key words： remote sensing images；object detection; geometric similarity；multipath feature fusion；ultra-fusion residual marching module

1引言

基于深度学习的方法在遥感图像目标检测领域已取得了显著进展[1]。（剩余22987字）