海军航空大学学报2026，Vol.41Issue(3)：547-556,10.DOI:10.7682/j.issn.2097-1427.2026.03.012
基于改进DeepLab V3+网络的无人机影像语义分割方法
Segmentation Method for UAV Imagery Based on Improved DeepLab V3+
摘要
Abstract
To address the challenges of poor semantic segmentation performance and blurred boundaries caused by sig-nificant differences in target scales and uneven sample distribution in UAV imagery,an improved DeepLab V3+seman-tic segmentation network is proposed.The proposed method includes the following enhancements:1)a CA-PA attention mechanism is designed to improve the model's ability to extract critical information and features from multi-scale tar-gets;2)an FFM feature fusion module is introduced to effectively integrate low-level and high-level features,thereby al-leviating boundary blurriness and enhancing the segmentation of small targets;and 3)a weighted hybrid loss function combining Dice Loss and Focal Loss is adopted to address class imbalance and improve the learning of challenging samples.Experimental results based on the Aeroscapes and UAVid datasets show that compared to existing semantic segmentation networks,the proposed method improves the processing ability of multi-scale targets and imbalanced samples,and achieves more accurate boundary segmentation,making it highly suitable for UAV imagery semantic seg-mentation tasks.关键词
语义分割/DeepLab V3+/注意力机制/特征融合/混合损失函数Key words
semantic segmentation/DeepLab V3+/attention mechanism/feature fusion/hybrid loss function分类
信息技术与安全科学引用本文复制引用
苏楠,赵振华..基于改进DeepLab V3+网络的无人机影像语义分割方法[J].海军航空大学学报,2026,41(3):547-556,10.基金项目
江苏省自然科学基金优秀青年基金(BK20230091) （BK20230091）
国家自然科学基金(62573230) （62573230）
