基于Conv-Involution的红外视频小样本人体行为识别方法  

Few-shot human behavior recognition method of infrared video based on Conv-Involution

作　　者：姚天 余磊[1] 崔帅华 周啸辉 熊邦书[1] 欧巧凤[1] YAO Tian;YU Lei;CUI Shuai-hua;ZHOU Xiao-hui;XIONG Bang-shu;OU Qiao-feng(Key Laboratory of Image Processing and Pattern Recognition of Jiangxi Province,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学图像处理与模式识别江西省重点实验室,江西南昌330063

出　　处：《激光与红外》2023年第2期246-252,共7页Laser & Infrared

基　　金：国家自然科学基金项目(No.62162044,No.61866027);江西省自然科学基金项目(No.20202BAB202016);江西省重点研发计划项目(No.20212BBE53017)资助。

摘　　要：红外视频存在颜色信息缺失较为严重、识别区域易与背景混淆等现象,使得现有小样本特征提取网络常常关注无效信息,导致识别精度较低。针对此问题,本文提出一种基于内卷积(Conv-Involution)算子的红外视频小样本人体行为识别方法。首先,通过InstColorization方法恢复红外视频中的颜色信息;其次,基于空间和通道特异性设计Conv-Involution算子,并利用该算子和改进注意力机制设计特征提取网络,分离背景,更有效地关注行为发生区域;最后,结合小样本学习方法ANIL进行人体行为分类。对比实验结果表明,本文所提方法不但识别精度最高,而且具有出色的稳定性。Infrared video suffers serious lack of color information, and the recognition area is easily confused with the background, which makes the existing few-shot feature extraction network frequently follow invalid data, resulting in low recognition accuracy.To address this problem, few-shot human behavior recognition of infrared video based on the Conv-Involution operator is proposed.Firstly, the color information in the infrared video is restored by the InstColorization method.Secondly, the Conv-Involution operator is designed based on spatial and channel specificity, and the feature extraction network is designed using this operator and an improved attention mechanism to separate the background and focus more effectively on the region where the behavior occurs.Finally, the human behavior is classified by combing the few-shot learning method.The comparative experimental results show that the proposed method not only has the highest recognition accuracy, but also has excellent stability.

关 键 词：人体行为识别 Conv-Involution算子 小样本学习 红外视频 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]