摘要
近年来,机器学习研究不断取得突破,促成了大量智能系统的成熟和落地。然而,当前“深度学习+大规模标注数据+完备先验知识”的机器学习范式过度依赖先验知识的完备性,其应用场景局限于静态封闭的专用系统。现实应用环境具有更多开放性和复杂性,例如现实环境中所包含的类别空间在训练期间无法被完全预知且会有新类别在测试阶段不断出现,这使得实际应用场景下的数据构成和分布都极其复杂,无法通过全局分析来保证模型的有效性。为了打破现有机器学习对完备类别信息的过度依赖,对开放集识别问题的研究已成为一个新的趋势。开放集识别将传统分类问题向开放环境下进行扩展,在保证已知类别准确分类的同时,要求模型还可以有效地识别测试阶段新出现的未知类别样本,避免造成大量误分。本文对近年来开放集识别的研究进行了系统调研,聚焦于基于深度学习的开放集识别方法,对经典模型进行了梳理和介绍,并对其分类效果进行了横向对比。
In recent years,machine learning research has continuously made breakthroughs,which has led to the maturity and implementation of a large number of intelligent systems.However,the current machine learning paradigm overly relies on complete prior knowledge,and its application scenarios are limited to static closed specialized systems.The emergence of new categories is one of the main challenges in the research of dynamic open environments,to which the conventional closed-world machine learning methods can not handle it well.In order to break the excessive dependence of existing machine learning on complete category information,the research on open set recognition problem has become a new trend,which extends the traditional closed-world classification methods to open-world applications.Open set recognition problem requires the classification models classify the learned known classes correctly,and effectively identify unknown class samples that appear in the testing phase simultaneously,to avoid the large number of misclassifications.This article systematically surveys the research on open set recognition in recent years,focusing on deep learning based open set recognition methods.It introduces the classic models by sorting out them into six categories based on the main ideas they employ,and horizontally compares various research achievements,and give the comparison details of their performances.Related paper and code links are collected and available online.
作者
章秦
刘紫琪
张晓林
张鹏
刘涵
陈小军
ZHANG Qin;LIU Zi-Qi;ZHANG Xiao-Lin;ZHANG Peng;LIU Han;CHEN Xiao-Jun(College of Computer Science and Software Engineering,ShenZhen University,ShenZhen,Guangdong 518060;College of Electrical Engineering and Automation,Shandong University of Science and Technology,Qingdao,Shandong 266590;College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao,Shandong 266590)
出处
《计算机学报》
北大核心
2025年第4期828-863,共36页
Chinese Journal of Computers
基金
国家自然科学基金项目(62206179,62202280,62106147)
广东省自然科学基金面上项目(2022A1515010129)
深圳市高等院校稳定支持计划(20220811121315001)
山东省自然科学基金项目(ZR2024QF034,ZR2021QF017)
山东省泰山学者工程青年专家项目(TSQN202312196)
深圳市科技计划项目(ZDSYS20220527171400002)资助。
关键词
开放集识别
深度学习
开放域
分类
open set recognition
deep learning
open world learning
classification
