用于近海区域植被保护的无人机图像语义分割算法研究
Research on UAV Image Semantic Segmentation Algorithm for Vegetation Protection in Offshore Areas
李庆华 1胥志伟 1赵天旭 1刘莹莹 2吴楷 2王胜科2
作者信息
- 1. 山东巍然智能科技有限公司
- 2. 中国海洋大学 信息科学与工程学部,山东 青岛 266100
- 折叠
摘要
海岸带对人类生活和经济发展有着深远影响,无人机在海洋生态保护中得到了广泛应用,但现有分割模型在无人机海岸带植被分割任务中还存在一些不足.一方面是无人机高度角度变化导致的零散小块植被分割困难和类内尺度变化大的问题,为此设计了CNN与Transformer 结合的特征提取网络,并由此设计了MEAFormer分支,以及类引导加权模块(CGW)来学习不同外观的鲁棒特征表示;另一方面是海岸带场景导致的相似植被类别分割错误与水下植被边界分割不清的问题,为此构建包含混合卷积注意力模块(MCA)与双注意力融合模块(DAFM)的融合分支来融合学习不同层次的特征,同时引入SAM分支,由MEAFormer分支得到Mask指导SAM进行精细化分割.实验结果表明,该方法在 Cityscapes 数据集和OUC-UAV-SEG 数据集上MIou分别达到了79.8%与72.4%,验证了该分割策略的有效性.
Abstract
Coastal zones have a profound impact on human life and economic development.UAV has been widely used in marine ecological protection.However,existing segmentation models still have some problems in UAV coastal zone vegetation segmentation tasks.Therefore,this paper designs a feature extraction network combined with CNN and Transformer,and then designs a MEAFormer branch.Meanwhile,a class-guided weighting module(CGW)is designed to learn the robust feature representation of different appearances.On the other hand,there are similar vegetation category segmentation errors and unclear underwater vegetation boundary segmentation caused by coastal zone scenes.Therefore,this paper constructs a fusion branch including mixed Convolution attention module(MCA)and dual attention fusion mod-ule(DAFM)to integrate and learn features of different levels.Meanwhile,SAM branch is introduced.The Mask obtained by the MEAFormer branch guides SAM to do fine segmentation.The MIou achieved 79.8%on Cityscapes and 72.4%on OUC-UAV-SEG,respectively,the effec-tiveness of the segmentation strategy proposed in this paper was verified.
关键词
植被保护/无人机图像分割/类引导加权/特征融合Key words
vegetation protection/unmanned aerial vehicle image segmentation/class guided weighting/feature fusion引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据