冗余度引导掩膜重构的小样本动作识别方法  

Redundancy Guided Masked Feature Reconstruction for Few-shot Action Recognition

作　　者：陈朋 颜灵强 党源杰 宦若虹[1] 梁荣华[1] CHEN Peng;YAN Lingqiang;DANG Yuanjie;HUAN Ruohong;LIANG Ronghua(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023

出　　处：《小型微型计算机系统》2024年第9期2188-2195,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(U1909203,62036009,62276237,62206250)资助;浙江省万人计划人才项目资助

摘　　要：目前大多数基于度量的小样本动作识别方法都采用支持样本与询问样本对比的方式学习视频特征,然而在样本量稀少的情况下,只使用对比的方法难以充分地挖掘视频中的信息.另一方面,视频中存在大量的冗余,有效地利用而不是单纯地减少这些冗余信息,可以提升小样本动作识别模型.为此,本文提出了一种冗余度引导特征掩膜重构的方法,将原始视频特征转换为若干离散的特征块,根据每个特征块的冗余程度排序并确定其是否被掩膜替换并重构,并且基于上述操作与原型网络实现小样本动作识别,使模型充分地利用低冗余的视频特征,实现模型泛化性能的提升.实验结果表明本文提出的冗余度引导掩膜重构的方法在5-way 5-shot的实验设置下在Kinetics和something-somethingv2数据集上均优于现有的小样本动作识别方法.Most of the current metric-based few-shot action recognition methods learn video features by comparing support samples with query samples.However,it is difficult to fully exploit the information in videos only by using comparison methods when the number of samples is scarce.On the other hand,there are a lot of redundancy in the video,effectively using redundant information instead of simply reducing it can improve the few-shot action recognition model.To this end,this paper proposes a redundancy-guided feature mask reconstruction method,which converts the original video features into several discrete feature blocks,sorts each feature block according to its redundancy metric then determines whether it is replaced by a mask and reconstructed,and based on the above operations and prototypical network to realize few-shot action recognition,so that the model can make full use of redundant video features and improve the generalization performance of the model.Experimental results show that our redundancy-guided mask reconstruction method proposed in this paper outperforms the existing few-shot action recognition methods.The experimental results show that the redundancy guided masked feature reconstruction method proposed in this paper is better than the existing few-shot action recognition methods on the Kinetics and something-somethingv2 dataset under the 5-way 5-shot experimental setting.

关 键 词：小样本动作识别 原型网络 元学习 特征重构 冗余度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]