| 摘要: |
| 针对遥感目标图像多样、目标较小且与周围环境对比度差，导致现有目标检测方法复杂度高、检测效果差、泛化能力弱，以及经典U-Net和U-SegNet忽略了不同目标尺度特征感受野差异等问题，提出了一种基于多尺度上下文注意力U-SegNet(Multi-scale Context Attention U-SegNet，MSCAUSNet)的遥感目标检测新模型。该模型由U-SegNet、多尺度特征融合(Multi-scale Feature Fusion,MSFF)和多尺度上下文注意力(Multi-scale Context Attention,MSCA)模块组成，采用MSCA代替U-SegNet中的跳跃连接以融合目标低层特征与高层特征，并通过MSFF和MSCA模块充分捕获多尺度上下文特征，从而显著提升遥感多尺度目标检测性能。在遥感目标图像数据集上的实验结果表明，该模型能够有效检测不同尺度遥感目标，较经典U-Net和U-SegNet的检测精度分别提高了7.94%和5.09%。该模型为遥感目标检测和识别系统提供了技术支持。 |
| 关键词: 遥感多尺度目标检测 多尺度上下文注意力U-SegNet 多尺度上下文注意力 多尺度特征融合 |
| DOI：10.20079/j.issn.1001-893x.241114005 |
|
| 基金项目:河南省科技攻关项目(232102210034，242102210021）；河南省教育厅项目(25A520049)；郑州西亚斯学院重点学科信息与通信工程(0810) |
|
| Remote Sensing Target Detection Based on Multi-scale Contextual Attention U-SegNet |
| CHEN Ruixia,ZHANG Shanwen,WU Qing揺 |
| (1.School of Telecommunications and Intelligent Manufacturing,Zhengzhou Sias University,Zhengzhou 451150,China;2.Henan Agricultural Information Digital Intelligence Engineering Research Center,Zhengzhou 451150,China;3.School of Computer and Information Engineering,Shanghai Polytechnic University,Shanghai 201209,China) |
| Abstract: |
| To solve the problems of high complexity,poor detection effect and weak generalization of existing remote sensing target detection models due to the variety of remote sensing target images,small target and low contrast with the surrounding environment,and the fact that U-Net and U-SegNet ignore the difference in the receptive field of different scale features,a multi-scale context attention U-SegNet(MSCAUSNet) is constructed.It consists of U-SegNet,multi-scale feature fusion(MSFF) and multi-scale context attention(MSCA) modules.In the model,MSCA instead of skip connections in U-SegNet is used to fuse low-level and high-level features of the target,and multi-scale context feature is captured by MSFF and MSCA modules,thus significantly improving the detection performance of remote sensing multiscale targets.The experimental results on the remote sensing target image dataset show that the proposed model is effective for remote sensing multiscale target detection.Compared with that of the original U-Net and U-SegNet,the detection rate of MSCAUSNet is increased by 7.94% and 5.09%,respectively.The proposed model provides a technical support for remote sensing target detection and recognition system. |
| Key words: remote sensing multi-scale target detection multi-scale context attention U-SegNet multi-scale context attention multi-scale feature fusion |
《电讯技术》编辑部 