基于元增量学习的开放集识别方法  

Open Set Recognition Based on Meta Class Incremental Learning

作　　者：孙晋永[1] 王雪纯 蔡国永[1] 尚之量 SUN Jinyong;WANG Xuechun;CAI Guoyong;SHANG Zhiliang(Guangxi Key Laboratory of Trusted Software,Guilin University of Electronic Technology,Guilin,Guangxi 541004,China)

机构地区：[1]桂林电子科技大学广西可信软件重点实验室,广西桂林541004

出　　处：《计算机科学》2025年第5期187-198,共12页Computer Science

基　　金：国家自然科学基金(62366010,61862016,62006058,62066010);广西可信软件重点实验室(KX202205);“认知无线电与信息处理”省部共建教育部重点实验室主任基金项目(CRKL210107)。

摘　　要：传统图像分类算法假定世界是静态、封闭的,而大数据时代的真实世界却是动态、开放的,新类别及其样本不断出现,导致传统图像分类算法的准确率降低。针对这种情况,研究者提出了适用于真实世界的开放集识别问题,目标是从样本集中识别出未知类样本,同时保持对已知类样本的分类准确性。但现有的开放集识别方法都忽略了对识别出的未知类样本的进一步利用,且未知类样本通常数量较少,这些情况导致开放集识别模型无法增量地学习到已识别出的未知类样本蕴含的知识,影响了开放集识别模型的准确性和泛化性。为此,提出一种基于元增量学习的开放集识别方法,来提高开放集识别模型的准确性和泛化性。该方法使用双层优化机制构建开放集识别模型,对未知类样本进行深度聚类,使模型能够对聚类后的未知类样本进行增量学习。具体来说,首先,构建基于双层优化机制的开放集识别模型,并对其进行训练,使其具备对少量未知类样本进行增量学习的能力。然后,使用权重激励注意力机制来获取开放集识别模型参数的重要性,对模型的非关键参数进行更新,减少增量学习对模型的已知类分类能力的影响。其次,设计深度DBSCAN方法对未知类样本进行聚类,将每簇样本标记为一类,并使模型对其增量学习,丢弃离散样本,减少离散样本对增量学习效果的影响。最后,在4个公开数据集上进行实验,结果表明,相较于主流的开放集识别方法,所提方法在AUROC和F1分数上均具有更好的效果,可以充分地学习识别出的未知类样本的知识。Traditional image classification algorithms assume that the world is static and closed,whereas the real world is dyna-mic and open,and new categories and their samples are continually emerging,leading to a decrease in the accuracies of traditional image classification algorithms.To address this problem,researchers proposed open set recognition(OSR)problem for the real world which aims at identifying unknown-class samples while maintaining the classification accuracy for known-class samples.However,existing OSR methods generally neglect the further exploitation of identified unknown-class samples and the unknown class samples are scarce in number,so that the classification model is unable to incrementally learn the knowledge of identified unknown class samples,thereby impairing the accuracy and generalization capability of OSR models.Therefore,this paper proposes an OSR method based on meta-incremental learning to improve the accuracy and generalization of OSR models.This method employs a bi-level optimization mechanism to build an OSR model,and then cluster unknown class samples based on deep learning so that the built OSR model can incrementally learn the knowledge of unknown class samples.Specifically,an OSR model based on bi-level optimization mechanism is constructed and trained with few-shot unknown class samples,in order to enable the OSR model to incrementally learn the knowledge of few-shot unknown class samples.Then,a weight excitation attention method is used to obtain the importance of the OSR model’s parameters and update non-critical parameters,thereby reducing the impact of incremental learning on the model’s ability to classify known-classes.Additionally,a deep learning-based DBSCAN method is designed to extract features and cluster the identified unknown-class samples.Clustered samples are labeled as the same class and performed incremental learning.While samples that are difficult to cluster are rejected,to avoid the impact of too few unknown-class samples on the model’s incremental learnin

关 键 词：开放集识别 图像分类 增量学习 元学习 聚类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]