[author_cn_name].[cn_title][J].空军工程大学学报,[year_id],[volume]([issue]):[start_page]-[end_page] 融合多尺度注意力和双向LSTM的行人重识别-A Pedestrian Re ID with Multi Scale Attention and Bidirectional LSTM
文章摘要
闫昊雷,李小春,张仁飞,张磊,邱浪波,王哲.融合多尺度注意力和双向LSTM的行人重识别[J].空军工程大学学报,2022,23(5):71-76
融合多尺度注意力和双向LSTM的行人重识别
A Pedestrian Re ID with Multi Scale Attention and Bidirectional LSTM
  
DOI:
中文关键词: 注意力机制;卷积神经网络;行人重识别  深度学习;LSTM
英文关键词: attention mechanism  convolutional neural network  pedestrian re ID  deep learning  LSTM
基金项目:陕西省重点发展计划(2021ZDLSF06 09)
作者单位
闫昊雷,李小春,张仁飞,张磊,邱浪波,王哲 1.空军工程大学信息与导航学院西安,7100772.武警陕西省总队西安,710054 3.陕西省信息化工程研究院西安,7100614.陆军装备部北京,100000 
摘要点击次数: 45
全文下载次数: 41
中文摘要:
      将深度学习应用于行人重识别领域,嵌入多尺度注意力融合模块至神经网络中进行多尺度特征提取和表示,可有效提升注意力机制对深度学习网络的识别性能。提出了一种基于SE block的多尺度通道注意力融合模块,并结合ResNet50卷积神经网络提取特征;然后通过双向LSTM网络进一步提取特征序列上下文信息,在提高模型对图像重要特征的提取能力的同时,降低对图像冗余特征的关注度;最后使用级联难采样三元组损失函数和交叉熵损失函数共同训练网络模型,使样本能够在高维特征空间中实现聚类,进一步提升模型识别准确性。所提出算法在Market1501数据集和CUHK03数据集分别进行实验,并在同等条件下和其他注意力模块算法进行比较。为进一步验证各模块作用,对算法进行消融实验,以验证各模块的有效性,实验结果表明,所提出方法可有效应用于行人重识别
英文摘要:
      With the rapid development of the information society, taking video sensors as the front end for acquiring information is of great significance in effectively finding specific objects through pedestrian re identification algorithms to protect people’s lives and property. This paper applies deep learning to the field of person re identification, and embeds the multi scale attention fusion module into the neural network for multi scale feature extraction and representation, effectively improving the recognition performance of the attention mechanism for deep learning networks. The paper proposes a multi scale channel attention fusion module based on SE block in combination with the ResNet50 convolutional neural network to extract features, further extract the feature sequence context information through the bidirectional LSTM network, and improve the model’s ability to extract important image features. At the same time, the attention to redundant features of images is reduced. Finally, the network model is jointly trained by the cascaded hard sampling triplet loss function and the cross entropy loss function, clustering the samples in the high dimensional feature space, and further improving the model recognition accuracy. Market1501 dataset and CUHK03 dataset are tested by the proposed algorithm respectively, and compared with other attention module algorithms under the same conditions. In order to further verify the function of each module, an ablation experiment is performed by the algorithm to verify the effectiveness of each module. The experimental results show that the proposed method can be effectively applied to person re identification
查看全文   查看/发表评论  下载PDF阅读器
关闭