文献类型：期刊文献

中文题名：基于GSegFormer的无人机图像密集小目标语义分割方法

英文题名：Semantic segmentation methods for dense small objects in UAV images based on GSegFormer

作者：刘炟[1,2];龙浩[1,2];张明瑜[2]

第一作者：刘炟

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100027

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2025

卷号：43

期号：4

起止页码：89-94

中文期刊名：飞行力学

外文期刊名：Flight Dynamics

收录：;北大核心:【北大核心2023】；

基金：国家重点研发计划资助(2022YFB4601104);北京联合大学校级科研项目资助(ZK20202304,ZKZD202302);北京联合大学“启明星”大学生科技创新创业项目资助(20242014)。

语种：中文

中文关键词：无人机;级联门控注意力;特征融合;语义分割

外文关键词：UAV;cascaded gated attention;feature fusion;semantic segmentation

摘要：针对现有模型经过多次下采样处理,遗漏小目标特征,对密集小目标分割效果有限,以及目标在不同距离下尺度差异大造成的分割能力不足等问题,提出了一种改进的基于门控注意力的语义转换器模型(GSegFormer)网络。设计了多尺度低损特征融合网络融合编码器提取的多尺度特征,包括串联门控注意力机制(CGAM)模块和多尺度目标特征融合(MTFF)模块。CGAM模块动态处理语义特征并抑制影响详细信息的特征,从而提高对小目标的分割精度。MTFF模块对4个不同尺度的特征上采样后进行拼接,在CGAM之前冻结权重参数,从而加速训练。在AeroScapes和ISPRS Vaihingen数据集上进行了试验,通过与主流方法的定量分析和比较表明,GSegFormer对密集小目标有更理想的分割性能,更适合于航拍影像小目标分割任务。
In response to the limitations of existing models,such as the omission of small object features due to multiple downsampling processes,limited segmentation effects on dense small objects,and insufficient segmentation capabilities caused by large target scale differences at different distances,an improved gated attention-based semantic transformer model(GSegFormer)network is proposed.A multi-scale low-loss feature fusion network is designed to fuse the multi-scale features extracted by the encoder,including a cascade gated attention mechanism(CGAM)module and a multi-scale target feature fusion(MTFF)module.The CGAM module dynamically processes semantic features and suppresses features that affect detailed information,thereby improving the segmentation accuracy of small targets.The MTFF module concatenates features of four different scales after the upsampling process and freezes the weight parameters before the CGAM to accelerate training.The proposed method has been tested on the AeroScapes and ISPRS Vaihingen datasets.Quantitative analysis and comparison with the mainstream methods indicate that GSegFormer has a more ideal segmentation performance for dense small objects and is more suitable for small object segmentation tasks on challenging aerial images.