文献类型：期刊文献

英文题名：Dual-level correspondence network for few-shot semantic segmentation

作者：Hui, Huang[1,2]; Wen, Chunlin[1]; Ma, Yan[1]; Yuan, Feiniu[1,2]; Zhu, Hongqing[3]; Zhu, Peng[4]

机构：[1] College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Haisi, Shanghai, 200234, China; [2] Shanghai Engineering Research Center of Intelligent Education and Bigdata, Shanghai Normal University, Haisi, Shanghai, 200234, China; [3] School of Information Science and Engineering, East China University of Science and Technology, Meilong, Shanghai, 200237, China; [4] Shanghai Vixdetect Inspection Equipment Co., Ltd., Jindu, Shanghai, 201108, China

年份：2025

卷号：84

期号：31

起止页码：37701

外文期刊名：Multimedia Tools and Applications

收录：EI(收录号：20251218080032)

语种：英文

外文关键词：Adversarial machine learning - Contrastive Learning - Query processing - Semantics - Zero-shot learning

摘要：Few-shot semantic segmentation (FSS) aims to segment novel class objects with only a few labeled support images from the same class. Most previous works exploit the prototype learning or affinity learning framework to extract single-level correspondence between support and query sets. However, single-level correspondence from the object or pixel information fails to fully mine semantic correlation, thus leading to incomplete segmentation or background noise. To address this issue, we propose the Dual-Level Correspondence Network (DLCNet) to establish the complementary correspondence with support prototype and pixel information guidance. The dual-level correspondence generation module accomplishes the dense matching between query features and dual-level object information to establish dual-level correspondence. Moreover, we design the attention mask generation module to alleviate the generalization reduction based on the multi-level prior attention and introduce the multi-scale feature adaptive fusion module to boost fusion features and refine fine-grained segmentation. Extensive experiments on PASCAL-5i and COCO-20i demonstrate the superiority of our method. ? The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2025.