摘要
针对现有的小样本语义分割方法普遍采用简单的融合策略来实现跨分支信息传递,难以捕获复杂的特征交互信息的问题,提出一个实现支持-查询分支高效融合的互引导条件耦合网络.所提网络设计了一个耦合信息生成模块,该模块聚焦于支持特征中与查询特征高度相似的区域,以使查询特征中的每个像素能够自适应地融合支持特征的像素级信息.同时,为了减少查询图像背景对融合效果的负面影响,设计了一个条件耦合模块.在该模块中,生成的条件可用于抑制查询图像背景区域的信息吸收,从而有效地避免了背景信息对结果的干扰.在PASCAL-5i以及COCO-20i的实验结果表明,所提网络在多数任务上,其mIoU和FBIoU指标优于或接近相关方法的最佳性能.
Existing few-shot semantic segmentation methods commonly employ simple fusion strategies to propagate cross-branch information,making it challenging to capture intricate feature interactions.To address this issue,a mutually guided conditional coupling network is proposed to efficiently fuse support-query branch information.The network incorporates a coupling information generation module,which focuses on regions in the support feature space highly similar to the query features,enabling each pixel in the query features to adaptively integrate pixel-level information from the support features.Additionally,to mitigate the adverse impact of querying image background on fusion effectiveness,a conditional coupling module is designed.In this module,the generated conditions are utilized to suppress information absorption in the query image background regions,effectively avoiding interference from background information on the results.Extensive experiments and analyses on PASCAL-5i and COCO-20i datasets demonstrate significant improvements compared to baseline methods.
作者
刘胜男
姜璐璐
郑宛露
王少荣
汪国平
Liu Shengnan;Jiang Lulu;Zheng Wanlu;Wang Shaorong;Wang Guoping(School of Information Science and Technology,Beijing Forest University,Beijing 100083;Beijing Virtual Simulation and Visualization Engineering Center,Beijing 100871;School of Computer Sciences,Peking University,Beijing 100871)
出处
《计算机辅助设计与图形学学报》
北大核心
2025年第6期973-982,共10页
Journal of Computer-Aided Design & Computer Graphics
关键词
语义分割
小样本学习
小样本语义分割
特征融合
semantic segmentation
few-shot learning
few-shot semantic segmentation
feature fusion
