嵌入标签语义的元特征再学习和重加权小样本目标检测  

Meta-Feature Relearning with Embedded Label Semantics and Reweighting for Few-Shot Object Detection

作　　者：李鹏芳 刘芳 李玲玲[1,2,3,4] 刘旭 冯志玺[1,2,3,4] 焦李成 熊怡梦[1,2,3,4] LI Peng-Fang;LIU Fang;LI Ling-Ling;LIU Xu;FENG Zhi-Xi;JIAO Li-Cheng;XIONG Yi-Meng(School of Artificial Intelligent,Xidian University,Xi’an 710071;Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education,Xi’an 710071;International Research Center for Intelligent Perception and Computation,Xi’an 710071;Joint International Research Laboratory of Intelligent Perception and Computation,Xi’an 710071)

机构地区：[1]西安电子科技大学人工智能学院,西安710071 [2]教育部智能感知与图像理解重点实验室,西安710071 [3]国际智能感知与计算研究中心,西安710071 [4]国际智能感知与计算联合研究实验室,西安710071

出　　处：《计算机学报》2022年第12期2561-2575,共15页Chinese Journal of Computers

基　　金：国家自然科学基金项目(62076192);陕西省重点研发计划(2019ZDLGY03-06);国家自然科学基金国家重点项目(61836009);长江学者及大学创新研究团队计划(IRT_15R53);高等学校学科创新引智计划(B07048);教育部重点科技创新研究项目;国家重点研发计划;CAAI华为MindSpore开放基金等资助.

摘　　要：小样本目标检测(Few-Shot Object Detection,FSOD)中新类相对基类样本少,且新类和基类目标类别不同,导致FSOD方法存在学习到的新类特征判别性不强的问题.为了增强新类元特征的可分性,本文提出了一种嵌入标签语义的元特征再学习和重加权小样本目标检测方法.在小样本训练阶段,本文构建了一个词向量标签语义图产生模块.该产生模块引入标签语义信息生成了词向量标签语义图,用于建模基类和新类间的语义关联.同时,本文构建了一个标签语义嵌入模块.该嵌入模块融入基类和新类间的语义关联,对支持集样本的元特征进行再学习.该再学习过程能够将基类中与新类相关联的特征传递给新类,从而在只有少量新类样本的情况下学习到较好的新类元特征.通过端到端(End-to-End)的训练模型,本文方法增强了新类元特征的可分性,从而提升了新类目标的检测精度.在PASCAL VOC和COCO数据集上的对比和消融实验表明了本文方法的可行性与有效性.与FSODFR方法相比,在PASCAL VOC数据集上2-shot和5-shot下,我们方法的目标检测精度分别提高了2.2%和4.3%.Object Detection methods based on Deep Learning(DL)have been able to achieve good detection accuracy.Nevertheless,DL,as a data-driven technique,relies heavily on massive labeled data.Currently,Few-Shot Learning has been extensively studied,as it can alleviate the reliance on a large number of labeled samples.In this paper,we focus on the research on Few-Shot Object Detection(FSOD).In FSOD,the deep model is first learned on the base class data that must have enough labeled samples.The model then continues to learn new classes with only a few labeled samples.The ultimate goal of model learning is to quickly adapt to the identification and localization of new classes of objects.In FSOD,on the one hand,compared with the base class,the new class has few samples available.On the other hand,the new class and the base class contain different target classes.This results in that FSOD methods are generally able to learn better base class features,but the learned new class features are weakly separable.To enhance the separability of meta-features for new classes,in this paper,we propose a Meta-Feature Relearning with Embedded Label Semantics and Reweighting Few-Shot Object Detection method.The proposed method can transfer the new class-related features from the base class to the new class,thereby enhancing the meta-features of the new class.In detail,in the few-shot training stage,firstly,we construct a word vector label semantic graph generation module.The word vector label semantic graph generation module introduces the label semantic information to generate the word vector label semantic graph.The generated word vector label semantic graph is used to model the semantic association between the base class and the new class.Meanwhile,we construct a label semantic embedding module.The label semantic embedding module first integrates the semantic association between the base class and the new class.Then,the meta-features of the support set samples are relearned based on the semantic association between the base class and the n

关 键 词：小样本学习 目标检测 小样本目标检测 元学习 标签语义 特征再学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]