燕山大学学报2025，Vol.49Issue(4)：300-308,9.DOI:10.3969/j.issn.1007-791X.2025.04.003
基于多原型交叉感知网络的小样本图像语义分割
Few-shot image semantic segmentation based on multi-prototype cross-perception network
摘要
Abstract
The information of support images alone is insufficient to provide sufficient guidance for segmenting unseen objects in the query image.To address this issue,a novel method for few-shot semantic segmentation based on the multi-prototype cross-perception network is proposed.Firstly,a set of shared weights backbone networks are used to map both support and query images into a deep feature space.In the support branch,the support feature map is decomposed into foreground and background feature maps using the ground true mask of the support image.Then,multiple prototype expressions are generated on the support foreground feature map using mask average pooling,and K-nearest neighbor clustering algorithm is used to generate multiple prototypes on the support background and query feature maps.Finally,the alignment of the two-branch prototype sets is achieved through cross-attention mechanisms,enhancing the perceptual ability of the prototype sets for the target task.Experimental results on the PASCAL-5 and COCO-20 datasets demonstrate that the proposed method achieves competitive segmentation performance on 1-shot and 5-shot tasks.关键词
小样本语义分割/交叉注意力机制/多原型/掩码平均池化/K近邻聚类算法Key words
few-shot semantic segmentation/cross-attention mechanism/multi-prototype/mask average pooling/K-nearest neighbor clustering algorithm分类
信息技术与安全科学引用本文复制引用
巴钧才,王昌龙..基于多原型交叉感知网络的小样本图像语义分割[J].燕山大学学报,2025,49(4):300-308,9.基金项目
国家自然科学基金资助项目(62362060),甘肃省教育厅创新基金项目(2022A-219) （62362060）
