全文下载： 
202402017.pdf
文章编号： 1672-6987（2024）02-0140-07； DOI： 10.16351/j.1672-6987.2024.02.017
王静， 王传旭*(青岛科技大学 信息科学技术学院， 山东 青岛 266061)
摘要： 弱监督行为检测旨在使用视频级标签定位动作的起止边界及识别相应的行为类别。现有的模型依然存在行为定位不完整、背景干扰等问题。对此，提出了双分支协同策略，为背景帧引入辅助类，采用权重共享机制的非对称式训练，使得该模型能够抑制背景帧的激活以提高定位性能。在优化分支提出中心损失项来学习每个动作类的聚类中心，并惩罚特征与其中心之间的距离及最小类内变化，从而增强特征的可辩别性；基本分支丢弃其动作类的中心区域，同时学习背景特征，通过迭代训练挖掘与其行为相关的不明显区域，有助于更好的模拟背景，实现行为的完整性定位。该算法在THUMOS14和ActivityNet1.2数据集上进行实验验证并与其他相关文献进行比较，结果表明了所提出算法的可行性。
关键词： 时序行为检测； 弱监督学习； 中心损失项； 背景类； 注意力机制
中图分类号： TP 391文献标志码： A
引用格式： 王静， 王传旭. 双分支协同策略的弱监督行为检测［J］. 青岛科技大学学报(自然科学版), 2024, 45(2): 140-146.
WANG Jing， WANG Chuanxu. Two-Branch collaborative strategy for weakly supervised behavior detection［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2024， 45（2）： 140-146.
Two-Branch Collaborative Strategy for Weakly Supervised Behavior Detection
WANG Jing， WANG Chuanxu
(College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China)
Abstract: Weakly supervised behavior detection aims to locate start and end boundaries of action and identify corresponding behavior categories using video-level labels. Most existing models still have problems such as incomplete behavior localization and background inference. In this regard, the paper proposes a two-branch collaborative strategy, which introduces auxiliary classes for background frames, and adopts asymmetric training with a weight sharing mechanism, so that the model can suppress the activation of background frames to improve localization performance. In the optimization, the center loss term is proposed, which aims to enhance their discriminability through action-specific clustering and minimizing the intra-class variations. The basic branch discards the central regions of the action classes and learns background features. Through iterative training, mining inconspicuous regions related to behaviors helps to better simulate the background and achieve complete positioning of behavior. Extensive experiments are carried out on THUMOS14 and ActivityNet1.2 datasets and compared with other relevant literature, which proves the feasibility of the proposed algorithm.
Key words: temporal behavior localization; weakly supervised learning; center loss term; background class; attention mechanism.
收稿日期： 2023-04-27
基金项目： 国家自然科学基金项目（61672305）.
作者简介： 王静（1998—），女，硕士研究生.*通信联系人.
