«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年10期

页码:

114-121

栏目:

人工智能

出版日期:

2025-10-10

文章信息/Info

Title:

Few-shot Semantic Segmentation Based on Adaptive Prototype Enhancement and Prior Fusion

文章编号:

1673-629X(2025)10-0114-08

作者:

王锦丽1; 2; 张智1; 2*

1. 武汉科技大学 计算机科学与技术学院,湖北武汉 430065;

2. 武汉科技大学 智能信息处理与实时工业系统湖北省重点实验室,湖北武汉 430065

Author(s):

WANG Jin-li1; 2; ZHANG Zhi1; 2*

1. School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;

2. Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan University of Science and Technology,Wuhan 430065,China

关键词:

小样本语义分割; 小样本学习; 元学习; 自适应原型增强; 先验知识

Keywords:

few-shot semantic segmentation; few-shot learning; meta-learning; adaptive prototype enhancement; prior knowledge

分类号:

TP391.41

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0125

摘要:

随着小样本语义分割任务的日益发展,如何在基类干扰较大的情况下提升模型的泛化能力和新类别的识别精度,成为了当下需要研究的问题。基于元学习框架的小样本语义分割在面对这些挑战时,存在泛化能力不足和在新类别识别时特征模糊的问题。为了有效解决这些问题,该文提出了一种基于自适应原型增强及先验融合的小样本语义分割算法,旨在显著提升模型在新类别上的分割性能,并增强其对未知类别的适应性。该文在BAM模型的基础上,引入自适应原型增强模块(PAM)从支持集提取的类别信息,派生出更加精确的类原型,从而增强模型的类别特定特征的表示能力;设计了先验掩码生成器(PMG)融合查询图像和支持图像的深度特征,提升目标前景区域的定位精度,并生成更鲁棒的先验掩码。在1-shot设置下,该方法在Pascal-5i和COCO-20i数据集上的性能分别提升了0.59%和0.46%;在5-shot设置下,两个数据集的性能提升进一步扩大,分别达到0.72%和0.55%的显著改进。实验结果表明,该方法能够有效地提升模型在新类别上的分割性能,具有较好的泛化能力和准确性。

Abstract:

With the growing development of few-shot semantic segmentation tasks,improving the model’s generalization ability and the recognition accuracy of new classes in the presence of large base class interference has become a hot topic in current research. Few-shot semantic segmentation based on meta-learning frameworks faces challenges such as insufficient generalization ability and blurred feature representations when recognizing new classes. To address these issues effectively,a novel few-shot semantic segmentation algorithm based on adaptive prototype enhancement and prior fusion is proposed,aiming to significantly improve the model’s segmentation performance on new classes and enhance its adaptability to unseen categories. In this work,an adaptive prototype enhancement module (PAM) is introduced on top of the BAM to derive more precise class prototypes from the category information extracted from the support set,thereby strengthening the model’s ability to represent class-specific features. Additionally,a prior mask generator (PMG) is designed to fuse deep features from both query and support images,improving the accuracy of foreground region localization and generating more robust prior masks. Experiments on the Pascal-5i and COCO-20i datasets show that in the 1-shot setting,the proposed method improves performance by 0.59% and 0.46%,respectively,and in the 5-shot setting,it improves by 0.72% and 0.55%. Experi-mental results demonstrate that the proposed method effectively enhances the model’s segmentation performance on new classes,with improved generalization capability and accuracy.

常用功能

更新日期/Last Update: 2025-10-10