首页|优化特征融合的多尺度遥感图像目标检测方法

优化特征融合的多尺度遥感图像目标检测方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
为解决遥感图像场景下多尺度目标检测准确率低的问题，提出DAFFNet遥感图像目标检测算法。该算法基于SSD进行了三方面的改进:为增强多尺度特征信息的获取能力，设计一种基于分组的特征融合方法;引入基于注意力机制的多维度特征优化方法，来解决复杂背景下目标分类困难的问题;将Focal loss作为新的边界框置信度损失函数，令模型聚焦于难分类的正样本，以改善正负样本不平衡对目标分类所造成的干扰。在遥感公共数据集NWPU VHR-10上进行模型评估，实验结果表明，该算法相较于原算法均值平均精度提高5。1百分点，能有效地提高遥感图像目标检测准确率。
MULTI-SCALE REMOTE SENSING IMAGE TARGET DETECTION METHOD BASED ON OPTIMIZED FEATURE FUSION
A DAFFNet remote sensing image object detection algorithm is proposed to solve the problem of low accuracy of multi-scale object detection in the remote sensing image scene.Based on SSD,the algorithm was improved in three aspects.We designed a group-based feature fusion method to enhance the ability to acquire multi-scale feature information.A multi-dimensional feature optimized method based on the attention mechanism was introduced to solve the difficulty of target classification in a complex background.The focal loss was used as a new bounding box confidential loss function to make the model focus on the positive samples that were difficult to classify,so as to improve the interference caused by the imbalance of positive and negative samples to target classification.The model was evaluated on the remote sensing public dataset NWPU VHR-10.The experimental result shows that the proposed algorithm improves the mean average precision by 5.1 percentage points compared with the original algorithm,which can effectively increase the object detection accuracy of remote sensing image.

Remote sensing imagesObject detectionGroup-based feature fusionMulti-dimensional feature optimizationAttention mechanism

张昊、刘凤、谭富祥、钱育蓉

展开 >

新疆大学软件学院 新疆乌鲁木齐 830008

新疆维吾尔自治区信号检测与处理重点实验室 新疆乌鲁木齐 830008

新疆大学软件工程重点实验室 新疆乌鲁木齐 830046

遥感图像 目标检测 分组特征融合 多维度特征优化 注意力机制

国家自然科学基金项目国家自然科学基金联合基金项目自治区科技厅国际合作项目智能多模态信息处理团队项目自治区研究生创新项目自治区研究生创新项目自治区研究生创新项目

61966035U18032612020E01023XJEDU2017T002XJ2019G069XJ2019G071XJ2020G074

2024

10.3969/j.issn.1000-386x.2024.08.022

计算机应用与软件
上海市计算技术研究所 上海计算机软件技术开发中心

计算机应用与软件

CSTPCD北大核心
影响因子：0.615
ISSN：1000-386X
年,卷(期)：2024.41(8)
  • 1