多尺度查询-支持注意力的小样本语义分割  

Few-shot Semantic Segmentation via Multi-scale Query-support Attention

作　　者：苏棋 吴宏亮 黄思恺 李洋 李东方 SU Qi;WU Hongliang;HUANG Sikai;LI Yang;LI Dongfang(College of Electrical Engineering and Automation,Fuzhou University,Fuzhou 350108,China;Ningdong Power Supply Company of State Grid Ningxia Electric Power Co.,Ltd.,Yinchuan750411,China)

机构地区：[1]福州大学电气工程与自动化学院,福州350108 [2]国网宁东供电公司,银川750411

出　　处：《无人系统技术》2025年第2期91-99,共9页Unmanned Systems Technology

基　　金：国家自然科学基金(62303117);福建省自然科学基金(2024J01278)。

摘　　要：小样本语义分割旨在通过少量标注样本实现新类别的精准分割,但现有方法因特征压缩导致的空间信息丢失、跨尺度对齐偏差及背景干扰等问题限制了其性能。为此,提出了一种基于多层多尺度查询-支持点积注意力的小样本语义分割框架。首先,支持图像和查询图像通过多尺度特征提取模块,从主干网络中提取多尺度特征。接着,利用Transformer的缩放点积注意力机制,充分挖掘支持图像中前景与背景像素的信息。最后,通过跳跃连接将多尺度特征与中间掩码进行融合,逐步细化分割边界。该模型在PASCAL-5i数据集上进行了实验,在1-shot设定下取得了62.1%的平均交并比和73.6%的前景-背景交并比,验证了方法的有效性。Few-shot semantic segmentation aims to achieve precise segmentation of novel classes with limited annotated samples,yet existing methods suffer from spatial information loss due to feature compression,cross-scale misalignment,and background interference.To address these challenges,this paper proposes a Multi-level Multi-scale Query-support Dot-product Attention(MLMS-DPA)framework.Firstly,the support image and query image extract multi-scale features from the backbone network through a multi-scale feature extraction module.Then,the scaled dot-product attention mechanism of the Transformer is utilized to fully exploit the information of foreground and background pixels in the support image.Finally,multi-scale features and intermediate masks are fused through skip connections to progressively refine the segmentation boundaries.The model was evaluated on the PASCAL-5i dataset,achieving a mean intersection-over-union of 62.1%and a foreground-background IoU of 73.6%under the 1-shot setting,demonstrating the effectiveness of the proposed method.

关 键 词：小样本 语义分割 缩放点积注意力 多层多尺度特征融合 Transformer架构 PASCAL-5i数据集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]